research article feature selection for intelligent...
TRANSCRIPT
Research ArticleFeature Selection for Intelligent Firefighting RobotClassification of Fire Smoke and Thermal Reflections UsingThermal Infrared Images
Jong-Hwan Kim1 Seongsik Jo1 and Brian Y Lattimer2
1Mechanical amp Systems Engineering Department Korea Military Academy Seoul Republic of Korea2Mechanical Engineering Department Virginia Tech Blacksburg VA 24060 USA
Correspondence should be addressed to Jong-Hwan Kim jonghwan7028gmailcom
Received 26 March 2016 Revised 5 July 2016 Accepted 3 August 2016
Academic Editor Juan A Corrales
Copyright copy 2016 Jong-Hwan Kim et alThis is an open access article distributed under theCreativeCommonsAttribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
Locating a fire inside of a structure that is not in the direct field of view of the robot has been researched for intelligent firefightingrobots By classifying fire smoke and their thermal reflections firefighting robots can assess local conditions decide a properheading and autonomously navigate toward a fire Long-wavelength infrared camera images were used to capture the scenedue to the camerarsquos ability to image through zero visibility smoke This paper analyzes motion and statistical texture featuresacquired from thermal images to discover the suitable features for accurate classification Bayesian classifier is implemented toprobabilistically classify multiple classes and a multiobjective genetic algorithm optimization is performed to investigate theappropriate combination of the features that have the lowest errors and the highest performance The distributions of multiplefeature combinations that have 670 or less error were analyzed and the best solution for the classification of fire and smoke wasidentified
1 Introduction
Intelligent firefighting humanoid robots are actively beingresearched to reduce firefighter injuries and deaths as wellas increase their effectiveness on performing tasks [1ndash5]One task is locating a fire inside of a structure outside therobot field of view (FOV) Fire smoke and their thermalreflections can be clues to determine a heading that willultimately lead the robot to the fire so that it can suppress itHowever research for accurately classifying these clues hasbeen incomplete
2 Previous Features
In conventional fire (andor smoke) detection systems [6 7]in Table 1 temperature ionization and ultraviolet light weremainly used to indicate the presence of a fire andor smokeinside the structure but they can have a long response timein large spaces [8] and do not provide sufficient data for thelocation of fire andor smoke Recently using vision systems
color [9ndash12] motion [13 14] both [8 15ndash17] and texture fea-tures [12 18 19] have been researched to characterize fireor smoke in Table 1 However color features from RGBcamera are not applicable to firefighting robots due to thefact that RGB cameras may operate in the visible to shortwavelength infrared (IR) (less than 1 micron) and are notusable in smoke-filled environments where the visibility hassufficiently decreased [2 14]Motion (eg dynamicalmotionshape changing etc) of the feature can be another clue todetect fire and smoke by characterizing flickering flames andsmoke flow from a stationary vision system However thevision system onboard a robot is moving due to the dynamicsof the robot itself and this causes a large amount of noise thatresults in extensive computation for motion compensationTexture features researched in [12 18 19] were used to identifyfire or smoke The spatial characteristics of textures can beuseful to recognize patterns of fire and smoke by remotesensing and are less influenced by rotationmotion [18]
Long-wavelength infrared cameras similar to the hand-held thermal infrared cameras (TICs) that are typically used
Hindawi Publishing CorporationJournal of SensorsVolume 2016 Article ID 8410731 13 pageshttpdxdoiorg10115520168410731
2 Journal of Sensors
Table 1 Conventional and vision-based features
Type Feature Advantages Disadvantages
Conventional features[6 7]
TemperatureIonizationUV light
(i) Detect presence of fireand smoke [8]
(i) Long response time [8](ii) Unable to providesufficient data for firelocating
Model-basedfeatures
Fourier transform [20]Wavelet transform [9]
(i) Frequency contentanalysis(ii) Flexible analysis of bothspace and frequency [25]
(i) Unable to be spatiallylocalized [25]
Vision-based features
Color (RGB) [9ndash12 26] (i) Fire (red)(ii) Smoke (gray)
(i) RGB camera cannotfunction in smoke-filledenvironments [2 14]
Dynamics [13 14] (motionshape change etc)
(i) Flickering flamesrecognition(ii) Smoke flow detection
(i) Can be influenced bydynamical robot motion(ii) Expensive computationfor motion compensation
Texture [12 18 19 27]
(i) Spatial characteristicsfor pattern recognition(ii) Less influenced byrotation and motion [18]
(i) The higher the ordertexture features the morethe computation
Feature maps [28] (CNNdeep learning)
(i) Superior performance inpattern recognition [29](ii) Once trainedapplicable in real-time
(i) Slow learning speed(ii) GPUs required due toexpensive computation
to aid in firefighting tasks within smoke-filled environments[20ndash22] as well as fire-front and burned-area recognition inremote sensing [23] are used in this research Due to the factthat TICs absorb infrared radiation in the long-wavelengthIR (7ndash14 microns) they are able to image surfaces even indense smoke and zero visibility environments [2 14] Inaddition TIC can provide proper information under localor global darkness for example shadows or darkness causedby damaged lighting Recently thermal images from TIC arestudied to recognize pattern and motion remotely [24] Thecameras will detect hot objects as well as thermal reflectionsoff of surfaces As a result image processing on detectedobjects must be sufficiently robust to discern between desiredobjects and their thermal reflections
This study ultimately leads the shipboard autonomousfirefighting robot (SAFFiR) whose prototype is displayed inFigure 1 to autonomously navigate toward fire outside FOVin indoor fire environments For this the robot needs toidentify clues such as smoke and smoke andfire-reflections byitself to correctly navigate toward the fireHowever the recog-nition of key features has not been fully studied This paperanalyzes appropriate combination of features to accuratelyclassify fire smoke their thermal reflections and other hotobjects using thermal infrared images Large-scale fire testswere conducted to create actual fire environments havingvarious ranges of both temperature and smoke conditionsA long-wavelength IR camera was installed to produce 14-bit thermal images of the fire environment These imageswere used to extract motion and statistical texture featuresin regions of interest (ROI) Bayesian classification wasperformed to probabilistically identify multiple classes inreal-time To identify the best combination of features for
Figure 1 A prototype of the shipboard autonomous firefightingrobot (SAFFiR) Note that the data used in this paper were notacquired from this platform
accurate classification the multiobjective optimization wasimplemented using two objective functions resubstitutionand cross-validation errors
3 Motion and Texture Features
In pattern recognition system the choice of features playsan important role in the performance of classification Both
Journal of Sensors 3
(a) RGB images
(b) Thermal images
Figure 2 (a) RGB images of fire scenes and (b) extracted objectsfrom thermal images with optical flow vectors overlaid
motion and texture features were selected because theywere crucial in the previous study of fire andor smokedetection and also best suitable for the thermal image analysisthat is major information the firefighting robot can acquireunder fire environments Optical flow a popular motionmeasurement was used for the motion features while thefirst- and second-statistical texture features were applied forthe texture measurement
A FLIR A35 long-wavelength IR camera which is capableof imaging through zero visibility environments was usedto produce images All images were from a 320 times 256-pixelfocal plane array 60Hz frame rate that produces 14-bit imageswith an intensity range of minus16384 for minus40∘C to minus1 for 550∘CFifteen features from optical flow and the statistical texturefeatures are evaluated to find the best feature combinationOptical flow shows temporal variations due tomoving objectsin the FOV or motion of the robot The first- and second-order statistical texture features display spatial characteristicsof objects in the scene
31 Motion Features by Optical Flow Optical flow is a usefultool to recognize motion of an object in sequential images[30] It consists of local and global methods Lucas-Kanade(LK) is a local method that is relatively robust with a lessdense flowfield whileHorn-Schunck (HS) is a globalmethodwith a dense flow field and high sensitivity to noise [31]Because the intensities in the thermal image change due tothe varying fire environment LK method that has higherrobustness compared with HS was selected in this researchto measure motion features of the objects Two featuresof optical flow vector number (OFVN) and optical flowmeanmagnitude (OFVMM)were computed to quantitativelycharacterize motions of fire smoke and their reflectionsFigure 2 contains RGB and thermal images of dense smokein a hallway and a wood crib fire in a room Red arrows in the
thermal images indicate the direction and magnitude of theoptical flow vectors with red boxes that show smoke fire andthermal reflections
32 First- and Second-Order Statistical Texture Features Thefirst- and second-order statistical features were considered inthis study for object classification The first-order statisticalfeatures estimate individual property of pixels not character-izing any relationship between neighboring pixels and canbe computed using the intensity histogram of the candidateregion of interest (ROI) in the image As described in [32]mean (MNI) variance (VAR) standard deviation (STD)skewness (SKE) and kurtosis (KUR) were calculated by
MNI = 1
119873119875
119873119875
sum
119894119895=1
119868119894119895
VAR = 1
119873119875
119873119875
sum
119894119895=1
(119868119894119895minus 120583)
2
STD (120590) = ( 1
119873119875
119873119875
sum
119894119895=1
(119868119894119895minus 120583)
2
)
12
SKE = 1
1205903
119873119875
119873119875
sum
119894119895=1
(119868119894119895minus 120583)
3
KUR = 1
1205904
119873119875
119873119875
sum
119894119895=1
(119868119894119895minus 120583)
4
(1)
where 119868119894119895
refers to the intensity of a pixel at 119894 and 119895 and119873119875denotes the number of pixels (NOP) of the object in the
image The second-order statistical features represent spatialrelationships between a pixel and its neighbors Gray-levelcooccurrence matrix (GLCM) [33] is used to account foradjacent pixel relationships in four directions (horizontalvertical left and right diagonals) by quantizing the spatialcooccurrence of neighboring pixels A total of seven second-order statistics features were used including dissimilarity(DIS) entropy (ENT) contrast (CON) inverse difference(INV) correlation (COR) uniformity (UNI) and inverse dif-ferencemoment (IDM) Tomeasure these features a normal-ized cooccurrence matrix 119862
119894119895is used which can be defined as
119862119894119895=
119875119894119895
sum119873119866
119894119895=1
119875119894119895
(2)
where 119875119894119895refers to the frequency of occurrences of the gray-
level of adjacent pixels at 119894 and 119895 within the four directionsand 119873
119866denotes the number of the gray-levels in the
quantized image The denominator of (2) normalizes 119875119894119895to
be estimates of the cooccurrence probabilities After buildingthe normalized cooccurrence matrix 119862
119894119895 seven features of
the second-order statistics features were computed by
DIS = sum119862119894119895
1003816100381610038161003816119894 minus 119895
1003816100381610038161003816
ENT = minussum119862119894119895log119862119894119895
4 Journal of Sensors
CON = sum119862119894119895(119894 minus 119895)
2
INV = sum119862119894119895
1 +1003816100381610038161003816119894 minus 119895
1003816100381610038161003816
COR = sum(119894119895 + 120583
119894120583119895minus 120583119894minus 120583119895) 119862119894119895
120590119894120590119895
UNI = sum1198622119894119895
IDM = sum
119862119894119895
1 + (119894 minus 119895)2
(3)
4 Object Extraction andBayesian Classification
One of the main characteristics of fire smoke and theirthermal reflections in thermal images is that they are higherin intensity than the background With intensity related totemperature in the thermal image higher temperature objectsappear brighter than the background Hence intensity is aprimary factor for object extraction from the backgroundAssuming that the thermal image histogram has a bimodaldistribution for foreground (ie object) and backgroundthe clustering-based image autothresholding method [34]called Otsumethod can calculate an optimum threshold thatseparates objects and background creating a binary imagewith 0 being the background and 1 being the objects Thebinary images were filtered to remove small regions and holesinside objects through morphological filtering techniquesAfter convoluting the original 14-bit image with the filtered-binary image a final image was obtained that includes theoriginal 14-bit intensities in objects as well as zeroes in thebackground
There are several classification methods commonly usedin supervised machine learning 119896-nearest neighbors (119896NN)decision tree (DT) neural networks (NN) support vectormachine (SVM) and Naıve Bayesian For this study theseclassification methods were analyzed by considering threepoints capability to classify multiple classes such as firesmoke and their thermal reflections less chance of overfit-ting problem because under fire environments there couldbe a number of situations that are not learned or trainedreal-time implementation because firefighting robot needs tomake a decision in real-time otherwise it cannot operate itstask 119896NN is insensitive to outliers but it needs a large amountof memory and expensive computation [35] DT has lowcomputation burden but for the multiclasses classificationit may generate a complicated tree structure and may causeoverfitting problem [35 36] NN shows high performancewhen processing with multidimensions and continuous fea-tures but cannot overcome overfitting problem SVM pro-vides fast computation and the highest accuracy but it cannotbe used for the multilabel classification because it producesbinary results [37] Naıve Bayesian classification is Bayesrsquotheorem-based probabilistic classification and is popular for
pattern recognition applications Although this method haslower accuracy compared with other classifiers and assumesthat each feature is independent it has fast computationrobustness to untrained cases and less chance of overfitting[35] In addition this classification has the capability ofprobabilistic decision making over multiple classes with fastcomputation for real-time implementation In this studyBayesian classification is used for evaluation of each feature
With several given features 1198651 1198652 119865
119902 (motion and
texture features) we can calculate the probability that oneclass 119862
ℎ(fire smoke thermal reflections etc) corresponds
to the candidate 119896 by using a conditional probability 119896119901(119862ℎ|
11986511198652sdot sdot sdot 119865119902) also known as the posterior probability By using
Bayesrsquo theorem it can be written with prior likelihood andevidence as shown in
119896
119901 (119862ℎ| 11986511198652sdot sdot sdot 119865119902)
=
119896
119901 (119862ℎ)119896
119891 (11986511198652sdot sdot sdot 119865119902| 119862ℎ)
sum119862ℎ
119896119901 (119862ℎ)119896
119891 (11986511198652sdot sdot sdot 119865119902| 119862ℎ)
(4)
where 119896119901(119862ℎ) is the prior probability meaning it represents
candidate 119896 probability to be 119862ℎand can be calculated by
number of samples of class 119862ℎdivided by the total number
of samples 119896119891(11986511198652sdot sdot sdot 119865119902| 119862ℎ) is the likelihood function
and the denominator of (4) is the evidence that plays asa normalizing constant by the summation of productionbetween the prior and likelihood at each class By applying theconditional independence assumption the likelihood func-tion can be rewritten by
119891 (11986511198652sdot sdot sdot 119865119902| 119862ℎ) =
119902
prod
119894=1
119891 (119865119894| 119862ℎ) (5)
The conditional probability density function 119891(119865119894| 119862ℎ) sim
119873(120583119865119894|119862ℎ
Σ119865119894|119862ℎ
) can be described as
119891 (119865119894| 119862ℎ) =
1
radic2120587Σ119865119894|119862ℎ
119890minus(12)(119865119894minus120583119865119894|119862ℎ
)
119879Σ
minus1
119865119894|119862ℎ(119865119894minus120583119865119894|119862ℎ
)
(6)
where
120583119865119894|119862ℎ
= 120583119865119894+ Σ119879
119862ℎ119865119894
Σminus1
119862ℎ119862ℎ
(119862ℎminus 120583119862ℎ)
Σ119865119894|119862ℎ
= Σ119865119894119865119894minus Σ119879
119862ℎ119865119894
Σminus1
119862ℎ119862ℎ
Σ119862ℎ119865119894
(7)
As shown in Table 2 Gaussian parameters for fifteen featureswith respect to smoke smoke thermal reflection fire andfire thermal reflection were estimated by using the maximumlikelihood estimation [38] Probability density distributionsfor the entire features are illustrated in Figure 3 With (5) theevidence and then the posterior probability of each class werecalculated By applying the maximum priority decision rulein (8) the Bayesian classification was used to predict the classand probability of each candidate in the scene
class = argmax119862ℎ
(119901 (119862ℎ| 11986511198652sdot sdot sdot 119865119902))
prob = max (119901 (119862ℎ| 11986511198652sdot sdot sdot 119865119902))
(8)
Journal of Sensors 5
Table 2 Gaussian parameters
Smoke Smoke-reflection Fire Fire-reflection120583 Σ 120583 Σ 120583 Σ 120583 Σ
MNI minus12665119864 + 04 69008119864 + 02 minus13383119864 + 04 35543119864 + 02 minus59714119864 + 03 16050119864 + 03 minus71399119864 + 03 46070119864 + 02
VAR 47578119864 + 05 44154119864 + 05 44012119864 + 04 41227119864 + 04 10501119864 + 07 22343119864 + 06 13300119864 + 06 88178119864 + 05
NOP 10733119864 + 03 42287119864 + 03 32220119864 + 01 13676119864 + 02 22174119864 + 02 17105119864 + 03 59575119864 + 01 23911119864 + 02
STD 62146119864 + 02 29931119864 + 02 18314119864 + 02 10236119864 + 02 32170119864 + 03 38930119864 + 02 10534119864 + 03 46992119864 + 02
SKE 11672119864 minus 01 58208119864 minus 01 minus94009119864 minus 02 62857119864 minus 01 22230119864 minus 02 52287119864 minus 01 22385119864 minus 01 71380119864 minus 01
KUR 30045119864 + 00 20213119864 + 00 36553119864 + 00 16131119864 + 00 22283119864 + 00 11517119864 + 00 34848119864 + 00 11912119864 + 00
OFVN 28832119864 + 04 14156119864 + 04 11848119864 + 03 87345119864 + 02 11257119864 + 04 13722119864 + 04 29978119864 + 03 22034119864 + 03
OFVMM 86830119864 + 01 56259119864 + 01 11687119864 + 02 79444119864 + 01 17598119864 + 02 29308119864 + 02 12641119864 + 02 67234119864 + 01
DIS 71791119864 minus 02 17966119864 minus 02 28045119864 minus 02 12245119864 minus 02 10937119864 minus 01 86515119864 minus 02 47308119864 minus 02 28550119864 minus 02
ENT 41949119864 minus 01 93683119864 minus 02 19576119864 minus 01 86106119864 minus 02 33681119864 minus 01 19107119864 minus 01 17115119864 minus 01 99702119864 minus 02
CON 89384119864 minus 01 36871119864 minus 01 12402119864 minus 01 65921119864 minus 02 92956119864 minus 01 69179119864 minus 01 39683119864 minus 01 26996119864 minus 01
IND 98588119864 minus 01 63845119864 minus 03 99646119864 minus 01 15141119864 minus 03 96334119864 minus 01 37231119864 minus 02 98771119864 minus 01 79275119864 minus 03
COR 96636119864 minus 01 31625119864 minus 02 88533119864 minus 01 43467119864 minus 02 89406119864 minus 01 30935119864 minus 02 88417119864 minus 01 57861119864 minus 02
UNI 51044119864 minus 01 20109119864 minus 01 95222119864 minus 01 31669119864 minus 02 67906119864 minus 01 26718119864 minus 01 85305119864 minus 01 10599119864 minus 01
IDM 60061119864 minus 01 19160119864 minus 01 97533119864 minus 01 16844119864 minus 02 77850119864 minus 01 24321119864 minus 01 92043119864 minus 01 58631119864 minus 02
Table 3 The object numbers of smoke smoke-reflection fire fire-reflection and other hot objects classes
Type Total Smoke Smoke-reflection Fire Fire-reflection Other hot objectsNumber of objects 10775 5190 1445 1464 489 2187
Figure 3 shows probability density distribution of eachclass using the Gaussian parameters of Table 2 Gaussiandistribution for classes in Figure 3 shows how fire fire-reflection smoke and smoke-reflection are distributed by thefifteen features Some features split out the distribution of thefour classes while others cause overlap For example MNIbest describes a well split out case of the classes althoughsmoke and its reflection and fire and its reflection do overlapSKE shows the worst case in which all classes overlap makingit impossible to distinguish any of the four classes
5 Result and Discussion
The accuracy in classifying fire objects was analyzed usingdata from a series of large-scale tests in the facility [1]using actual fires up to 75 kW Fires included latex foamwood cribs and propane gas fires from a sand burner Thesedifferent types of fires produced a range of temperature andsmoke conditions Latex foam fires produced lower tempera-ture conditions but dense low visibility smoke Converselypropane gas fires produced higher gas temperatures andlight smoke Wood crib fires resulted in smoke and gastemperatures between those of latex foam and propane gasfires however these fires resulted in sparks created from theburning wood Thermal images were collected by driving awheeled mobile robot through the setup during a fire test Atotal of 10775 objects were collected from the experimentsand categorized as either smoke smoke-reflection fire fire-reflection or other hot objects in order to be served as cluesto lead the firefighting robot to navigate toward the fire source
outside the FOV In addition as each object has sixteen cor-responding data points (fifteen features and a class) the totalnumber of data points used in this paper is 172400The num-bers of each object in this experiment are shown in Table 3
Two types of error criterions (resubstitution and 119896-foldcross-validation errors [39]) were used to measure howeach feature accurately performs in the classification Resub-stitution error takes the entire dataset to compare the actualclasses with the predicted classes by the Bayesian classifica-tion in order to examine how well the actual and predictedclasses match each other When this criterion is used aloneto enhance accuracy the classification can be overfitted tothe training dataset Cross-validation error is advantageousto detect and prevent from overfitting Instead of using theentire dataset cross-validation randomly selects and splitsthe dataset into 119896 partitions of approximately equal size(119896 = 10) to estimate a mean error by comparing betweenthe randomly selected partition and trained results of theremaining partitions
51 Single Feature Performance The performance results ofeach feature are shown in Table 4 The first-order statisticaltexture features MNI VAR and STD produced the lowesterrors while NOP SKE and KUR show the highest Theseresults show that MNI and VAR are beneficial to distinguishfire smoke and thermal reflections while motion features arenot As NOP shows the highest error OFVMM one of themotion features shows the second highest errors comparedwith the other features This is in part attributed to thedynamic motion of the robot ENT and COR second-order
6 Journal of Sensors
0 5 10VAR
Prob
abili
ty d
ensit
y
SmokeS-reflection
FireF-reflection
times10minus4
times10minus4
times104
times10minus7
times106
times10minus3
2468
1012
Prob
abili
ty d
ensit
y
minus10000 minus8000 minus6000 minus4000minus12000MNI
0
5
10
Prob
abili
ty d
ensit
y
0
05
1
Prob
abili
ty d
ensit
y
5000 100000OFVN
0
20
40
Prob
abili
ty d
ensit
y
01 02 03 040DIS
500 1000 1500 20000OFVMM
0
0005
001
Prob
abili
ty d
ensit
y0
5
Prob
abili
ty d
ensit
y
2 4 60NOP
50100150200250
Prob
abili
ty d
ensit
y
09 095 1085IND
1 15 2 2505CON
2
4
6
Prob
abili
ty d
ensit
y
02 04 06 08 10ENT
0
5
Prob
abili
ty d
ensit
y
2468
1012
Prob
abili
ty d
ensit
y
06 08 104COR
minus2 0 2 4 6minus4SKE
0
05
1
Prob
abili
ty d
ensit
y
01
02
03
Prob
abili
ty d
ensit
y
5 10 15 200KUT
1000 2000 3000 40000STD
0
2
4
08 10604IDM
510152025
Prob
abili
ty d
ensit
y
02 04 06 08 10UNI
0
10
20
Prob
abili
ty d
ensit
y
SmokeS-reflection
FireF-reflection
SmokeS-reflection
FireF-reflection
times10minus3
Figure 3 Probability density distributions of each feature
statistical texture features show42sim45error which is higherthan the other second-order features
52 Multiple Feature Combination Performance The errorresults in Table 4 demonstrate that a single feature cannotaccurately classify fire smoke and thermal reflections Thuspossible combination ofmultiple features was considered andanalyzed to find the best combination of the featuresThe totalnumber of all possible combinations that have two or morefeatures is
119873total =119898
sum
119911=2
(
119898
119911
) (9)
where 119898 refers to the total number of features (ie 119898 = 15)and 119911 is the number of features in the combination Basedon all possible combination the multiobjective genetic algo-rithm optimization [40] in the global optimization toolbox ofMATLAB was used to find the best combination of featuresthat has the highest performance in the classification Theobjective functions in the optimization resubstitution and119896-fold cross-validation errors [39] were used to measurehow accurately different feature combinations perform in theclassification
Figure 4 contains a plot of the error associated with themost promising feature combinations The behavioral solu-tion set is defined as feature combinations with less than 7
Journal of Sensors 7
Table 4 Performance of each feature
Resubstitution error () Cross-validation error ()MNI 237 238VAR 243 243NOP 721 720STD 232 232SKE 527 528KUR 506 506OFVN 395 400OFVMM 586 586DIS 290 290ENT 446 446CON 285 285IND 301 301COR 412 411UNI 345 345IDM 377 377
Behavioral region
Behavioral solution setGeneral set
66 67 68 69 7 71 72 7365Resubstitute error ()
65
66
67
68
69
7
71
72
73
Cros
s-va
lidat
ion
erro
r (
)
Figure 4Multiobjective optimization result showing the general setand behavioral solution set (colored region)
error for both objective functions while the general set refersto all other possible feature combinations The behavioralsolution set contains 00061 of all possible feature combi-nations
The occurrence probability of features in the behavioralsolution set is illustrated in Figure 5 In the behavioralsolution set the first-order statistic texture features MNI andSKE always exist while OFVN NOP and OFVMM featuresdo not Both the first-order statistical texture features STDand VAR and the second-order statistical texture featuresCOR ENT andDIS showahigher occurrence comparedwiththe other first- and second-order texture features while KURIDMUNI IND andCON show lower occurrence Note thatdue to the robotrsquos dynamical motion motion features werenot successful and even not included in the top 10 featurecombinations of the behavioral set
1000800
00900
1000200
0000
800900
100200
900200
400
MNIVARNOPSTDSKE
KUROFVN
OFMMDIS
ENTCONINDCORUNIIDM
200 400 600 800 1000 120000Occurrence probability ()
Figure 5 Occurrence analysis of the features in the behavioralregion
The top features based on the probability occurrence inFigure 5 are COR ENT DIS SKE STD VAR and MNIHowever the combination of these seven features does notresult in the best solution for classification Table 5 containsthe classification performance of the combination of featuresin the behavioral solution set In order to evaluate the per-formance of each feature combination various performancemeasures have been used such as precision sensitivity F-measure and accuracy Precision measures the fraction ofpositive instances from the group that the classifier predictedto be positive and recall measures the fraction of positiveexamples from the positive group of the actual class and [35]F-measure is the harmonic mean and accuracy is the pro-portion of true resultsThesemeasures can bemathematicallydefined as
Precision = TPTP + FP
Sensitivity = TPTP + FN
119865-Measure =2 (Sensitivity sdot Precision)Sensitivity + Precision
Accuracy = TPTP + FP + FN
(10)
where TP is correctly classified positive cases FP is incor-rectly classified negative cases and FN is incorrectly classifiedpositive cases For the performance measurement confusionmatrixes were created as described in Appendix and appliedinto (10) In the precision index number 1 combinationshows the highest performance in the behavioral solutionset while index number 7 combination shows the lowestIn the sensitivity index number 7 combination records thehighest results while index number 4 does the lowest In theF-measure and accuracy index number 2 combination showsthe highest record while index number 4 does the lowestBased on the confusion matrixes most of misclassificationoccurs in the classification of smoke smoke-reflection andother hot objects because during small fire texture patternsof these classes were diminished and the intensity was too low
8 Journal of Sensors
Table 5 Results of error case and performance at each feature combination in the behavioral solution set (Resu means resubstitution andCros refers to cross-validation)
Index Combination of features Error () Case Performance ()Resu Cros TP FP FN Precision Sensitivity 119865-measure Accuracy
1 MNI VAR ENT COR SKE 690 689 10049 402 324 9615 9688 9651 93262 MNI DIS COR SKE STD 668 670 10069 459 247 9564 9761 9662 93453 MNI ENT COR SKE STD 676 675 10061 447 267 9575 9741 9657 9337
4 MNI VAR DIS ENT CON CORSKE STD 698 690 9980 454 341 9565 9670 9617 9262
5 MNI VAR DIS ENT COR UNISKE STD 689 690 10037 453 285 9568 9724 9645 9315
6 MNI VAR DIS ENT COR IDMSKE STD 677 679 10047 461 267 9561 9741 9650 9324
7 MNI VAR DIS ENT IDM SKEKUR STD 696 694 10033 505 237 9521 9769 9643 9311
8 MNI VAR DIS ENT IND CORUNI SKE STD 695 699 10029 451 295 9570 9714 9641 9308
9 MNI VAR DIS ENT IND CORIDM SKE STD 688 690 10035 457 283 9564 9726 9644 9313
10 MNI VAR DIS ENT COR IDMSKE KUR STD 689 691 10036 478 261 9545 9747 9645 9314
PrecisionSensitivity
F-measureAccuracy
90919293949596979899
Perfo
rman
ce (
)
2 3 4 5 6 7 8 9 101Feature combination index
Figure 6 Results of performance at each feature combination in thebehavioral solution set
to distinguish The best solution was determined to be indexnumber 2 combination of MNI DIS COR SKE and STDwhich has the lowest of resubstitution and cross-validationerrors 668 and 670 respectively This combinationincludes all of the top features based on the probability occur-rence except ENT and VAR The four performance results ateach feature combination in the behavioral solution set areshown in Figure 6 where the highest results aremarked in redcircles and the lowest in green-dot circles Sensitivity appearshigher than precision at each feature combination becauseFPs are larger than FNs in the confusion matrix Particularlyindex number 7 has the biggest difference between FP and FNresulting in the highest sensitivity and lowest precision Thesummation of FP and FN in index number 4 is the highestin the behavioral solution set resulting in the lowest accuracywhile index number 2 has the lowest summation of FP andFN providing the highest accuracy
This study investigated a wide range of features fromlong-wavelength infrared camera images analyzed normaldistributions of fifteen features with respect to the classes ofsmoke fire and their thermal reflections and discovered thehighest performing feature combination by examining singlefeatures and multiple feature combinations As a result theproposed feature combination of MNI DIS COR SKE andSTD increases the performance compared with the previousstudy [1] which used MNI VAR ENT and IDM As shownin Figure 7 the errors are reduced by 286 and 268 resub-stitution and cross-validation errors and performances areincreased by 290 158 020 and 285 accuracy F-measure sensitivity and precision respectively
Figure 8 shows original visual and thermal images withthe robot at three different locations start point hallwayentrance and room entrance described in the experimentalfacility Each row relates to a series of images from the robotat three locations The first row contains visible images ofthe robot view As seen in the visible image at start pointfurther information regarding the hallway is limited due toshadowing of the light The image at hallway entrance showsa smoke layer in the upper portion of the hallway due to afire inside the roomThe image at the room entrance displaysa wood crib fire with sparks Because of soot and relativedifference in brightness the background is shown darker andthus limiting information on the background around the fire
Thermal infrared images are displayed in the second rowto show information that RGB camera cannot provide infire environments Unlike visual image at start point thatis obscured due to shadowing the presence of smoke andits thermal reflections on the ventilation hood can be obvi-ously perceived The red boxes on thermal images indicateobjects extracted through the adaptive object extraction withoptical flows and identification numbers In spite of dense
Journal of Sensors 9
9279
9741
9504
9055
954
938
9564
9761
9662
9345
668
670
Precision
Sensitivity
F-measure
Accuracy
Resubstitution error
Cross-validation error
The proposedThe previous
2000 4000 6000 8000 10000000()
Figure 7 Result comparison between the previous and the proposed studies
① Start point ② Hallway entrance ③ Room entrance
IR im
age
class
ifica
tion
resu
lts
Visib
le im
age
IR im
age
imag
e ana
lysis
Figure 8 Original visible and IR images with image analysis and classification results at different locations in the test setup of actual fires
smoke-filled and low visibility environments thermal imagescan generate the images of smoke and fire as well as back-ground information that is otherwise not visible throughvisual imaging
On the third row class labels and posterior probabilitiesof each candidate are displayed at the center of candidateROI as a result of Bayesian classification Using enhancedimage processing techniques the thermal images can bemore
10 Journal of Sensors
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
2
0
19
4820
64
Feature combination index number 2
Feature combination index number 3 Feature combination index number 4
Feature combination index number 5
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1462
0
0
0
1
1462
0
0
0
1
1457
0
0
0
1
1457
0
0
1365
193
197
0
0
1343
181
196
0
0
1325
205
181
0
0
1270
207
181
0
0
53
167
1929
0
0
145
166
1943
0
0
101
165
1942
7
0
49
4844
62
2
0
30
4817
63
0
0
53
165
1929
7
0
27
4830
61
MNI DIS COR SKE STD
MNI ENT COR SKE STD MNI VAR DIS ENT CON COR SKE STD
MNI VAR DIS ENT COR UNI SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bjPredicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Feature combination index number 1
140
159
0
0
0
0
0
00
0
0
0
0
0
1981
488
1280
193
144
1
1462
4838
62
25
2
MNI VAR ENT COR SKE
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Feature combination index number 6
0
0
0
0
488
0
0
0
1
14620
0
1339
206
185
0
0
93
161
1935
2
0
13
4823
67
MNI VAR DIS ENT COR IDM SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
F-re
flect
ion
Oth
er o
bj
Smok
e
Fire
S-re
flect
ion
Predicted class
Figure 9 Continued
Journal of Sensors 11
Feature combination index number 7 Feature combination index number 8
Feature combination index number 9 Feature combination index number 10
89
133
0
0
0
0
0
02
0
0
0
0
0
1888
488
0
0
0
0
488
1341
200
235
1
1459
4857
62
15
5
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1458
0
0
0
1
1462
0
0
0
1
1462
0
0
1324
204
181
0
0
1334
208
181
0
0
1334
205
204
0
0
99
174
1943
0
0
91
150
1921
2
0
18
4810
65
6
0
20
4835
62
0
0
93
172
1941
2
0
22
4812
63
MNI VAR DIS ENT COR IDM SKE KUR STD
MNI VAR DIS ENT IDM SKE KUR STD
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bj
Predicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
MNI VAR DIS ENT IND COR IDM SKE STD
MNI VAR DIS ENT IND COR UNI SKE STD
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
Figure 9 Results of confusion matrix at each feature combination
refined and clearer than the thermal images on the secondrow Smoke fire and their thermal reflections are identifiedand marked in red or orange ellipses
6 Conclusion
The appropriate combination of features was investigated toaccurately classify fire smoke and their thermal reflectionsusing thermal images Gray-scale 14-bit images from a singleinfrared camera were used to extract motion and texture fea-tures by applying a clustering-based autothresholding tech-nique Bayesian classification is performed to probabilisti-cally identify multiple classes during real-time implemen-tation To find the best combination of features a multi-objective genetic algorithm optimization was implementedusing resubstitution and cross-validation errors as objectivefunctions Large-scale fire tests with different fire sources
were conducted to create a range of temperature and smokeconditions to evaluate the feature combinations
Fifteenmotion and texture features were analyzed and theprobability density functions of the features were computedby the maximum likelihood estimation The combinationof multiple features was determined to more accuratelyclassify fire smoke and thermal reflections compared witha single feature In the behavioral solution set where featurecombinations produce less than 7 resubstitution and cross-validation errors COR ENT DIS SKE STD VAR andMNIhad 800 or more occurrence while other features had400 or less occurrence The feature combination of MNIDIS COR SKE and STD produced the highest performancein the classification resulting in 668 and 670 resubstitu-tion and cross-validation errors and 9564 9761 9662and 9345 precision sensitivity F-measure and accuracyrespectively
12 Journal of Sensors
In the near future the classification of fire smoke andtheir thermal reflections will be evaluated on any classifiersand features to increase performanceThe convolution neuralnetwork of deep learning which has recently shown highperformance could be explored as a classifier also model-based image features such as discrete wavelet transform willbe further studied
Appendix
See Figure 9
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this manuscript
Acknowledgments
This work was sponsored by the Office of Naval ResearchGrant no N00014-11-1-0074 scientific office Dr ThomasMcKenna in USA Hwarang-dae Research Institute in Seouland Agency for Defense Development in Daejeon SouthKorea The authors would like to thank Mr Joseph Starr andMr JoshMcNeil for assisting in performing the fire testsTheauthors would also like to thank Rosana K Lee for helpingand supporting this research
References
[1] J-H Kim and B Y Lattimer ldquoReal-time probabilistic classifi-cation of fire and smoke using thermal imagery for intelligentfirefighting robotrdquo Fire Safety Journal vol 72 pp 40ndash49 2015
[2] J W Starr and B Y Lattimer ldquoEvaluation of navigation sensorsin fire smoke environmentsrdquo Fire Technology vol 50 no 6 pp1459ndash1481 2014
[3] J-H Kim B Keller and B Y Lattimer ldquoSensor fusion basedseek-and-find fire algorithm for intelligent firefighting robotrdquoin Proceedings of the IEEEASME International Conference onAdvanced Intelligent Mechatronics (AIM rsquo13) pp 1482ndash1486IEEE Wollongong Australia July 2013
[4] J G McNeil J Starr and B Y Lattimer ldquoAutonomous fire sup-pression using multispectral sensorsrdquo in Proceedings of theIEEEASME International Conference on Advanced IntelligentMechatronics Mechatronics for HumanWellbeing (AIM rsquo13) pp1504ndash1509 Wollongong Austrailia July 2013
[5] J-H Kim J W Starr and B Y Lattimer ldquoFirefighting robotstereo infrared vision and radar sensor fusion for imagingthrough smokerdquo Fire Technology vol 51 no 4 pp 823ndash8452015
[6] RC Luo andK L Su ldquoAutonomous fire-detection systemusingadaptive sensory fusion for intelligent security robotrdquo IEEEASME Transactions on Mechatronics vol 12 no 3 pp 274ndash2812007
[7] M A Jackson and I Robins ldquoGas sensing for fire detectionmeasurements of CO CO
2
H2
O2
and smoke density in Euro-pean standard fire testsrdquo Fire Safety Journal vol 22 no 2 pp181ndash205 1994
[8] B U Toreyin R G Cinbis Y Dedeoglu and A E Cetin ldquoFiredetection in infrared video using wavelet analysisrdquo OpticalEngineering vol 46 no 6 Article ID 067204 2007
[9] B U Toreyin Y Dedeoglu andA E Cetin ldquoWavelet based real-time smoke detection in videordquo in Proceedings of the 13th Euro-pean Signal Processing Conference pp 4ndash8 Antalya TurkeySeptember 2005
[10] T Celik H Demirel H Ozkaramanli and M Uyguroglu ldquoFiredetection using statistical color model in video sequencesrdquoJournal of Visual Communication and Image Representation vol18 no 2 pp 176ndash185 2007
[11] L Merino F Caballero J R Martınez-de-Dios I Maza and AOllero ldquoAn unmanned aircraft system for automatic forest firemonitoring and measurementrdquo Journal of Intelligent amp RoboticSystems vol 65 no 1 pp 533ndash548 2012
[12] Y Wang T W Chua R Chang and N T Pham ldquoReal-timesmoke detection using texture and color featuresrdquo in Proceed-ings of the 21st International Conference on Pattern Recognition(ICPR rsquo12) pp 1727ndash1730 Tsukuba Japan November 2012
[13] G Marbach M Loepfe and T Brupbacher ldquoAn image process-ing technique for fire detection in video imagesrdquo Fire SafetyJournal vol 41 no 4 pp 285ndash289 2006
[14] M I Chacon-Murguia and F J Perez-Vargas ldquoThermal videoanalysis for fire detection using shape regularity and intensitysaturation featuresrdquo in Pattern Recognition J F Martınez-Trinidad J A Carrasco-Ochoa C B-Y Brants and E RHancock Eds vol 6718 of Lecture Notes in Computer Sciencepp 118ndash126 Springer Berlin Germany 2011
[15] W Phillips III M Shah and N D V Lobo ldquoFlame recognitionin videordquo Pattern Recognition Letters vol 23 no 1ndash3 pp 319ndash327 2002
[16] D Han and B Lee ldquoDevelopment of early tunnel fire detectionalgorithm using the image processingrdquo in Advances in VisualComputing pp 39ndash48 Springer Berlin Germany 2006
[17] Y Chunyu F Jun W Jinjun and Z Yongming ldquoVideo firesmoke detection using motion and color featuresrdquo Fire Technol-ogy vol 46 no 3 pp 651ndash663 2010
[18] F Yuan ldquoVideo-based smoke detection with histogram se-quence of LBP and LBPV pyramidsrdquo Fire Safety Journal vol 46no 3 pp 132ndash139 2011
[19] F Lafarge X Descombes and J Zerubia ldquoTextural kernel forSVM classification in remote sensing application to forest firedetection andUrban area extractionrdquo in Proceedings of the IEEEInternational Conference on Image Processing (ICIP rsquo05) pp1096ndash1099 September 2005
[20] F Amon and A Ducharme ldquoImage frequency analysis for test-ing of fire service thermal imaging camerasrdquo Fire Technologyvol 45 no 3 pp 313ndash322 2009
[21] F Amon V Benetis J Kim and A Hamins ldquoDevelopmentof a performance evaluation facility for fire fighting thermalimagersrdquo in Defense and Security pp 244ndash252 2004
[22] F D Maxwell ldquoA portable IR system for observing fire thrusmokerdquo Fire Technology vol 7 no 4 pp 321ndash331 1971
[23] A Barducci D Guzzi P Marcoionni and I Pippi ldquoInfrareddetection of active fires and burnt areas theory and observa-tionsrdquo Infrared Physics amp Technology vol 43 no 3ndash5 pp 119ndash125 2002
[24] C Wang and S Qin ldquoAdaptive detection method of infraredsmall target based on target-background separation via robustprincipal component analysisrdquo Infrared Physics amp Technologyvol 69 pp 123ndash135 2015
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
2 Journal of Sensors
Table 1 Conventional and vision-based features
Type Feature Advantages Disadvantages
Conventional features[6 7]
TemperatureIonizationUV light
(i) Detect presence of fireand smoke [8]
(i) Long response time [8](ii) Unable to providesufficient data for firelocating
Model-basedfeatures
Fourier transform [20]Wavelet transform [9]
(i) Frequency contentanalysis(ii) Flexible analysis of bothspace and frequency [25]
(i) Unable to be spatiallylocalized [25]
Vision-based features
Color (RGB) [9ndash12 26] (i) Fire (red)(ii) Smoke (gray)
(i) RGB camera cannotfunction in smoke-filledenvironments [2 14]
Dynamics [13 14] (motionshape change etc)
(i) Flickering flamesrecognition(ii) Smoke flow detection
(i) Can be influenced bydynamical robot motion(ii) Expensive computationfor motion compensation
Texture [12 18 19 27]
(i) Spatial characteristicsfor pattern recognition(ii) Less influenced byrotation and motion [18]
(i) The higher the ordertexture features the morethe computation
Feature maps [28] (CNNdeep learning)
(i) Superior performance inpattern recognition [29](ii) Once trainedapplicable in real-time
(i) Slow learning speed(ii) GPUs required due toexpensive computation
to aid in firefighting tasks within smoke-filled environments[20ndash22] as well as fire-front and burned-area recognition inremote sensing [23] are used in this research Due to the factthat TICs absorb infrared radiation in the long-wavelengthIR (7ndash14 microns) they are able to image surfaces even indense smoke and zero visibility environments [2 14] Inaddition TIC can provide proper information under localor global darkness for example shadows or darkness causedby damaged lighting Recently thermal images from TIC arestudied to recognize pattern and motion remotely [24] Thecameras will detect hot objects as well as thermal reflectionsoff of surfaces As a result image processing on detectedobjects must be sufficiently robust to discern between desiredobjects and their thermal reflections
This study ultimately leads the shipboard autonomousfirefighting robot (SAFFiR) whose prototype is displayed inFigure 1 to autonomously navigate toward fire outside FOVin indoor fire environments For this the robot needs toidentify clues such as smoke and smoke andfire-reflections byitself to correctly navigate toward the fireHowever the recog-nition of key features has not been fully studied This paperanalyzes appropriate combination of features to accuratelyclassify fire smoke their thermal reflections and other hotobjects using thermal infrared images Large-scale fire testswere conducted to create actual fire environments havingvarious ranges of both temperature and smoke conditionsA long-wavelength IR camera was installed to produce 14-bit thermal images of the fire environment These imageswere used to extract motion and statistical texture featuresin regions of interest (ROI) Bayesian classification wasperformed to probabilistically identify multiple classes inreal-time To identify the best combination of features for
Figure 1 A prototype of the shipboard autonomous firefightingrobot (SAFFiR) Note that the data used in this paper were notacquired from this platform
accurate classification the multiobjective optimization wasimplemented using two objective functions resubstitutionand cross-validation errors
3 Motion and Texture Features
In pattern recognition system the choice of features playsan important role in the performance of classification Both
Journal of Sensors 3
(a) RGB images
(b) Thermal images
Figure 2 (a) RGB images of fire scenes and (b) extracted objectsfrom thermal images with optical flow vectors overlaid
motion and texture features were selected because theywere crucial in the previous study of fire andor smokedetection and also best suitable for the thermal image analysisthat is major information the firefighting robot can acquireunder fire environments Optical flow a popular motionmeasurement was used for the motion features while thefirst- and second-statistical texture features were applied forthe texture measurement
A FLIR A35 long-wavelength IR camera which is capableof imaging through zero visibility environments was usedto produce images All images were from a 320 times 256-pixelfocal plane array 60Hz frame rate that produces 14-bit imageswith an intensity range of minus16384 for minus40∘C to minus1 for 550∘CFifteen features from optical flow and the statistical texturefeatures are evaluated to find the best feature combinationOptical flow shows temporal variations due tomoving objectsin the FOV or motion of the robot The first- and second-order statistical texture features display spatial characteristicsof objects in the scene
31 Motion Features by Optical Flow Optical flow is a usefultool to recognize motion of an object in sequential images[30] It consists of local and global methods Lucas-Kanade(LK) is a local method that is relatively robust with a lessdense flowfield whileHorn-Schunck (HS) is a globalmethodwith a dense flow field and high sensitivity to noise [31]Because the intensities in the thermal image change due tothe varying fire environment LK method that has higherrobustness compared with HS was selected in this researchto measure motion features of the objects Two featuresof optical flow vector number (OFVN) and optical flowmeanmagnitude (OFVMM)were computed to quantitativelycharacterize motions of fire smoke and their reflectionsFigure 2 contains RGB and thermal images of dense smokein a hallway and a wood crib fire in a room Red arrows in the
thermal images indicate the direction and magnitude of theoptical flow vectors with red boxes that show smoke fire andthermal reflections
32 First- and Second-Order Statistical Texture Features Thefirst- and second-order statistical features were considered inthis study for object classification The first-order statisticalfeatures estimate individual property of pixels not character-izing any relationship between neighboring pixels and canbe computed using the intensity histogram of the candidateregion of interest (ROI) in the image As described in [32]mean (MNI) variance (VAR) standard deviation (STD)skewness (SKE) and kurtosis (KUR) were calculated by
MNI = 1
119873119875
119873119875
sum
119894119895=1
119868119894119895
VAR = 1
119873119875
119873119875
sum
119894119895=1
(119868119894119895minus 120583)
2
STD (120590) = ( 1
119873119875
119873119875
sum
119894119895=1
(119868119894119895minus 120583)
2
)
12
SKE = 1
1205903
119873119875
119873119875
sum
119894119895=1
(119868119894119895minus 120583)
3
KUR = 1
1205904
119873119875
119873119875
sum
119894119895=1
(119868119894119895minus 120583)
4
(1)
where 119868119894119895
refers to the intensity of a pixel at 119894 and 119895 and119873119875denotes the number of pixels (NOP) of the object in the
image The second-order statistical features represent spatialrelationships between a pixel and its neighbors Gray-levelcooccurrence matrix (GLCM) [33] is used to account foradjacent pixel relationships in four directions (horizontalvertical left and right diagonals) by quantizing the spatialcooccurrence of neighboring pixels A total of seven second-order statistics features were used including dissimilarity(DIS) entropy (ENT) contrast (CON) inverse difference(INV) correlation (COR) uniformity (UNI) and inverse dif-ferencemoment (IDM) Tomeasure these features a normal-ized cooccurrence matrix 119862
119894119895is used which can be defined as
119862119894119895=
119875119894119895
sum119873119866
119894119895=1
119875119894119895
(2)
where 119875119894119895refers to the frequency of occurrences of the gray-
level of adjacent pixels at 119894 and 119895 within the four directionsand 119873
119866denotes the number of the gray-levels in the
quantized image The denominator of (2) normalizes 119875119894119895to
be estimates of the cooccurrence probabilities After buildingthe normalized cooccurrence matrix 119862
119894119895 seven features of
the second-order statistics features were computed by
DIS = sum119862119894119895
1003816100381610038161003816119894 minus 119895
1003816100381610038161003816
ENT = minussum119862119894119895log119862119894119895
4 Journal of Sensors
CON = sum119862119894119895(119894 minus 119895)
2
INV = sum119862119894119895
1 +1003816100381610038161003816119894 minus 119895
1003816100381610038161003816
COR = sum(119894119895 + 120583
119894120583119895minus 120583119894minus 120583119895) 119862119894119895
120590119894120590119895
UNI = sum1198622119894119895
IDM = sum
119862119894119895
1 + (119894 minus 119895)2
(3)
4 Object Extraction andBayesian Classification
One of the main characteristics of fire smoke and theirthermal reflections in thermal images is that they are higherin intensity than the background With intensity related totemperature in the thermal image higher temperature objectsappear brighter than the background Hence intensity is aprimary factor for object extraction from the backgroundAssuming that the thermal image histogram has a bimodaldistribution for foreground (ie object) and backgroundthe clustering-based image autothresholding method [34]called Otsumethod can calculate an optimum threshold thatseparates objects and background creating a binary imagewith 0 being the background and 1 being the objects Thebinary images were filtered to remove small regions and holesinside objects through morphological filtering techniquesAfter convoluting the original 14-bit image with the filtered-binary image a final image was obtained that includes theoriginal 14-bit intensities in objects as well as zeroes in thebackground
There are several classification methods commonly usedin supervised machine learning 119896-nearest neighbors (119896NN)decision tree (DT) neural networks (NN) support vectormachine (SVM) and Naıve Bayesian For this study theseclassification methods were analyzed by considering threepoints capability to classify multiple classes such as firesmoke and their thermal reflections less chance of overfit-ting problem because under fire environments there couldbe a number of situations that are not learned or trainedreal-time implementation because firefighting robot needs tomake a decision in real-time otherwise it cannot operate itstask 119896NN is insensitive to outliers but it needs a large amountof memory and expensive computation [35] DT has lowcomputation burden but for the multiclasses classificationit may generate a complicated tree structure and may causeoverfitting problem [35 36] NN shows high performancewhen processing with multidimensions and continuous fea-tures but cannot overcome overfitting problem SVM pro-vides fast computation and the highest accuracy but it cannotbe used for the multilabel classification because it producesbinary results [37] Naıve Bayesian classification is Bayesrsquotheorem-based probabilistic classification and is popular for
pattern recognition applications Although this method haslower accuracy compared with other classifiers and assumesthat each feature is independent it has fast computationrobustness to untrained cases and less chance of overfitting[35] In addition this classification has the capability ofprobabilistic decision making over multiple classes with fastcomputation for real-time implementation In this studyBayesian classification is used for evaluation of each feature
With several given features 1198651 1198652 119865
119902 (motion and
texture features) we can calculate the probability that oneclass 119862
ℎ(fire smoke thermal reflections etc) corresponds
to the candidate 119896 by using a conditional probability 119896119901(119862ℎ|
11986511198652sdot sdot sdot 119865119902) also known as the posterior probability By using
Bayesrsquo theorem it can be written with prior likelihood andevidence as shown in
119896
119901 (119862ℎ| 11986511198652sdot sdot sdot 119865119902)
=
119896
119901 (119862ℎ)119896
119891 (11986511198652sdot sdot sdot 119865119902| 119862ℎ)
sum119862ℎ
119896119901 (119862ℎ)119896
119891 (11986511198652sdot sdot sdot 119865119902| 119862ℎ)
(4)
where 119896119901(119862ℎ) is the prior probability meaning it represents
candidate 119896 probability to be 119862ℎand can be calculated by
number of samples of class 119862ℎdivided by the total number
of samples 119896119891(11986511198652sdot sdot sdot 119865119902| 119862ℎ) is the likelihood function
and the denominator of (4) is the evidence that plays asa normalizing constant by the summation of productionbetween the prior and likelihood at each class By applying theconditional independence assumption the likelihood func-tion can be rewritten by
119891 (11986511198652sdot sdot sdot 119865119902| 119862ℎ) =
119902
prod
119894=1
119891 (119865119894| 119862ℎ) (5)
The conditional probability density function 119891(119865119894| 119862ℎ) sim
119873(120583119865119894|119862ℎ
Σ119865119894|119862ℎ
) can be described as
119891 (119865119894| 119862ℎ) =
1
radic2120587Σ119865119894|119862ℎ
119890minus(12)(119865119894minus120583119865119894|119862ℎ
)
119879Σ
minus1
119865119894|119862ℎ(119865119894minus120583119865119894|119862ℎ
)
(6)
where
120583119865119894|119862ℎ
= 120583119865119894+ Σ119879
119862ℎ119865119894
Σminus1
119862ℎ119862ℎ
(119862ℎminus 120583119862ℎ)
Σ119865119894|119862ℎ
= Σ119865119894119865119894minus Σ119879
119862ℎ119865119894
Σminus1
119862ℎ119862ℎ
Σ119862ℎ119865119894
(7)
As shown in Table 2 Gaussian parameters for fifteen featureswith respect to smoke smoke thermal reflection fire andfire thermal reflection were estimated by using the maximumlikelihood estimation [38] Probability density distributionsfor the entire features are illustrated in Figure 3 With (5) theevidence and then the posterior probability of each class werecalculated By applying the maximum priority decision rulein (8) the Bayesian classification was used to predict the classand probability of each candidate in the scene
class = argmax119862ℎ
(119901 (119862ℎ| 11986511198652sdot sdot sdot 119865119902))
prob = max (119901 (119862ℎ| 11986511198652sdot sdot sdot 119865119902))
(8)
Journal of Sensors 5
Table 2 Gaussian parameters
Smoke Smoke-reflection Fire Fire-reflection120583 Σ 120583 Σ 120583 Σ 120583 Σ
MNI minus12665119864 + 04 69008119864 + 02 minus13383119864 + 04 35543119864 + 02 minus59714119864 + 03 16050119864 + 03 minus71399119864 + 03 46070119864 + 02
VAR 47578119864 + 05 44154119864 + 05 44012119864 + 04 41227119864 + 04 10501119864 + 07 22343119864 + 06 13300119864 + 06 88178119864 + 05
NOP 10733119864 + 03 42287119864 + 03 32220119864 + 01 13676119864 + 02 22174119864 + 02 17105119864 + 03 59575119864 + 01 23911119864 + 02
STD 62146119864 + 02 29931119864 + 02 18314119864 + 02 10236119864 + 02 32170119864 + 03 38930119864 + 02 10534119864 + 03 46992119864 + 02
SKE 11672119864 minus 01 58208119864 minus 01 minus94009119864 minus 02 62857119864 minus 01 22230119864 minus 02 52287119864 minus 01 22385119864 minus 01 71380119864 minus 01
KUR 30045119864 + 00 20213119864 + 00 36553119864 + 00 16131119864 + 00 22283119864 + 00 11517119864 + 00 34848119864 + 00 11912119864 + 00
OFVN 28832119864 + 04 14156119864 + 04 11848119864 + 03 87345119864 + 02 11257119864 + 04 13722119864 + 04 29978119864 + 03 22034119864 + 03
OFVMM 86830119864 + 01 56259119864 + 01 11687119864 + 02 79444119864 + 01 17598119864 + 02 29308119864 + 02 12641119864 + 02 67234119864 + 01
DIS 71791119864 minus 02 17966119864 minus 02 28045119864 minus 02 12245119864 minus 02 10937119864 minus 01 86515119864 minus 02 47308119864 minus 02 28550119864 minus 02
ENT 41949119864 minus 01 93683119864 minus 02 19576119864 minus 01 86106119864 minus 02 33681119864 minus 01 19107119864 minus 01 17115119864 minus 01 99702119864 minus 02
CON 89384119864 minus 01 36871119864 minus 01 12402119864 minus 01 65921119864 minus 02 92956119864 minus 01 69179119864 minus 01 39683119864 minus 01 26996119864 minus 01
IND 98588119864 minus 01 63845119864 minus 03 99646119864 minus 01 15141119864 minus 03 96334119864 minus 01 37231119864 minus 02 98771119864 minus 01 79275119864 minus 03
COR 96636119864 minus 01 31625119864 minus 02 88533119864 minus 01 43467119864 minus 02 89406119864 minus 01 30935119864 minus 02 88417119864 minus 01 57861119864 minus 02
UNI 51044119864 minus 01 20109119864 minus 01 95222119864 minus 01 31669119864 minus 02 67906119864 minus 01 26718119864 minus 01 85305119864 minus 01 10599119864 minus 01
IDM 60061119864 minus 01 19160119864 minus 01 97533119864 minus 01 16844119864 minus 02 77850119864 minus 01 24321119864 minus 01 92043119864 minus 01 58631119864 minus 02
Table 3 The object numbers of smoke smoke-reflection fire fire-reflection and other hot objects classes
Type Total Smoke Smoke-reflection Fire Fire-reflection Other hot objectsNumber of objects 10775 5190 1445 1464 489 2187
Figure 3 shows probability density distribution of eachclass using the Gaussian parameters of Table 2 Gaussiandistribution for classes in Figure 3 shows how fire fire-reflection smoke and smoke-reflection are distributed by thefifteen features Some features split out the distribution of thefour classes while others cause overlap For example MNIbest describes a well split out case of the classes althoughsmoke and its reflection and fire and its reflection do overlapSKE shows the worst case in which all classes overlap makingit impossible to distinguish any of the four classes
5 Result and Discussion
The accuracy in classifying fire objects was analyzed usingdata from a series of large-scale tests in the facility [1]using actual fires up to 75 kW Fires included latex foamwood cribs and propane gas fires from a sand burner Thesedifferent types of fires produced a range of temperature andsmoke conditions Latex foam fires produced lower tempera-ture conditions but dense low visibility smoke Converselypropane gas fires produced higher gas temperatures andlight smoke Wood crib fires resulted in smoke and gastemperatures between those of latex foam and propane gasfires however these fires resulted in sparks created from theburning wood Thermal images were collected by driving awheeled mobile robot through the setup during a fire test Atotal of 10775 objects were collected from the experimentsand categorized as either smoke smoke-reflection fire fire-reflection or other hot objects in order to be served as cluesto lead the firefighting robot to navigate toward the fire source
outside the FOV In addition as each object has sixteen cor-responding data points (fifteen features and a class) the totalnumber of data points used in this paper is 172400The num-bers of each object in this experiment are shown in Table 3
Two types of error criterions (resubstitution and 119896-foldcross-validation errors [39]) were used to measure howeach feature accurately performs in the classification Resub-stitution error takes the entire dataset to compare the actualclasses with the predicted classes by the Bayesian classifica-tion in order to examine how well the actual and predictedclasses match each other When this criterion is used aloneto enhance accuracy the classification can be overfitted tothe training dataset Cross-validation error is advantageousto detect and prevent from overfitting Instead of using theentire dataset cross-validation randomly selects and splitsthe dataset into 119896 partitions of approximately equal size(119896 = 10) to estimate a mean error by comparing betweenthe randomly selected partition and trained results of theremaining partitions
51 Single Feature Performance The performance results ofeach feature are shown in Table 4 The first-order statisticaltexture features MNI VAR and STD produced the lowesterrors while NOP SKE and KUR show the highest Theseresults show that MNI and VAR are beneficial to distinguishfire smoke and thermal reflections while motion features arenot As NOP shows the highest error OFVMM one of themotion features shows the second highest errors comparedwith the other features This is in part attributed to thedynamic motion of the robot ENT and COR second-order
6 Journal of Sensors
0 5 10VAR
Prob
abili
ty d
ensit
y
SmokeS-reflection
FireF-reflection
times10minus4
times10minus4
times104
times10minus7
times106
times10minus3
2468
1012
Prob
abili
ty d
ensit
y
minus10000 minus8000 minus6000 minus4000minus12000MNI
0
5
10
Prob
abili
ty d
ensit
y
0
05
1
Prob
abili
ty d
ensit
y
5000 100000OFVN
0
20
40
Prob
abili
ty d
ensit
y
01 02 03 040DIS
500 1000 1500 20000OFVMM
0
0005
001
Prob
abili
ty d
ensit
y0
5
Prob
abili
ty d
ensit
y
2 4 60NOP
50100150200250
Prob
abili
ty d
ensit
y
09 095 1085IND
1 15 2 2505CON
2
4
6
Prob
abili
ty d
ensit
y
02 04 06 08 10ENT
0
5
Prob
abili
ty d
ensit
y
2468
1012
Prob
abili
ty d
ensit
y
06 08 104COR
minus2 0 2 4 6minus4SKE
0
05
1
Prob
abili
ty d
ensit
y
01
02
03
Prob
abili
ty d
ensit
y
5 10 15 200KUT
1000 2000 3000 40000STD
0
2
4
08 10604IDM
510152025
Prob
abili
ty d
ensit
y
02 04 06 08 10UNI
0
10
20
Prob
abili
ty d
ensit
y
SmokeS-reflection
FireF-reflection
SmokeS-reflection
FireF-reflection
times10minus3
Figure 3 Probability density distributions of each feature
statistical texture features show42sim45error which is higherthan the other second-order features
52 Multiple Feature Combination Performance The errorresults in Table 4 demonstrate that a single feature cannotaccurately classify fire smoke and thermal reflections Thuspossible combination ofmultiple features was considered andanalyzed to find the best combination of the featuresThe totalnumber of all possible combinations that have two or morefeatures is
119873total =119898
sum
119911=2
(
119898
119911
) (9)
where 119898 refers to the total number of features (ie 119898 = 15)and 119911 is the number of features in the combination Basedon all possible combination the multiobjective genetic algo-rithm optimization [40] in the global optimization toolbox ofMATLAB was used to find the best combination of featuresthat has the highest performance in the classification Theobjective functions in the optimization resubstitution and119896-fold cross-validation errors [39] were used to measurehow accurately different feature combinations perform in theclassification
Figure 4 contains a plot of the error associated with themost promising feature combinations The behavioral solu-tion set is defined as feature combinations with less than 7
Journal of Sensors 7
Table 4 Performance of each feature
Resubstitution error () Cross-validation error ()MNI 237 238VAR 243 243NOP 721 720STD 232 232SKE 527 528KUR 506 506OFVN 395 400OFVMM 586 586DIS 290 290ENT 446 446CON 285 285IND 301 301COR 412 411UNI 345 345IDM 377 377
Behavioral region
Behavioral solution setGeneral set
66 67 68 69 7 71 72 7365Resubstitute error ()
65
66
67
68
69
7
71
72
73
Cros
s-va
lidat
ion
erro
r (
)
Figure 4Multiobjective optimization result showing the general setand behavioral solution set (colored region)
error for both objective functions while the general set refersto all other possible feature combinations The behavioralsolution set contains 00061 of all possible feature combi-nations
The occurrence probability of features in the behavioralsolution set is illustrated in Figure 5 In the behavioralsolution set the first-order statistic texture features MNI andSKE always exist while OFVN NOP and OFVMM featuresdo not Both the first-order statistical texture features STDand VAR and the second-order statistical texture featuresCOR ENT andDIS showahigher occurrence comparedwiththe other first- and second-order texture features while KURIDMUNI IND andCON show lower occurrence Note thatdue to the robotrsquos dynamical motion motion features werenot successful and even not included in the top 10 featurecombinations of the behavioral set
1000800
00900
1000200
0000
800900
100200
900200
400
MNIVARNOPSTDSKE
KUROFVN
OFMMDIS
ENTCONINDCORUNIIDM
200 400 600 800 1000 120000Occurrence probability ()
Figure 5 Occurrence analysis of the features in the behavioralregion
The top features based on the probability occurrence inFigure 5 are COR ENT DIS SKE STD VAR and MNIHowever the combination of these seven features does notresult in the best solution for classification Table 5 containsthe classification performance of the combination of featuresin the behavioral solution set In order to evaluate the per-formance of each feature combination various performancemeasures have been used such as precision sensitivity F-measure and accuracy Precision measures the fraction ofpositive instances from the group that the classifier predictedto be positive and recall measures the fraction of positiveexamples from the positive group of the actual class and [35]F-measure is the harmonic mean and accuracy is the pro-portion of true resultsThesemeasures can bemathematicallydefined as
Precision = TPTP + FP
Sensitivity = TPTP + FN
119865-Measure =2 (Sensitivity sdot Precision)Sensitivity + Precision
Accuracy = TPTP + FP + FN
(10)
where TP is correctly classified positive cases FP is incor-rectly classified negative cases and FN is incorrectly classifiedpositive cases For the performance measurement confusionmatrixes were created as described in Appendix and appliedinto (10) In the precision index number 1 combinationshows the highest performance in the behavioral solutionset while index number 7 combination shows the lowestIn the sensitivity index number 7 combination records thehighest results while index number 4 does the lowest In theF-measure and accuracy index number 2 combination showsthe highest record while index number 4 does the lowestBased on the confusion matrixes most of misclassificationoccurs in the classification of smoke smoke-reflection andother hot objects because during small fire texture patternsof these classes were diminished and the intensity was too low
8 Journal of Sensors
Table 5 Results of error case and performance at each feature combination in the behavioral solution set (Resu means resubstitution andCros refers to cross-validation)
Index Combination of features Error () Case Performance ()Resu Cros TP FP FN Precision Sensitivity 119865-measure Accuracy
1 MNI VAR ENT COR SKE 690 689 10049 402 324 9615 9688 9651 93262 MNI DIS COR SKE STD 668 670 10069 459 247 9564 9761 9662 93453 MNI ENT COR SKE STD 676 675 10061 447 267 9575 9741 9657 9337
4 MNI VAR DIS ENT CON CORSKE STD 698 690 9980 454 341 9565 9670 9617 9262
5 MNI VAR DIS ENT COR UNISKE STD 689 690 10037 453 285 9568 9724 9645 9315
6 MNI VAR DIS ENT COR IDMSKE STD 677 679 10047 461 267 9561 9741 9650 9324
7 MNI VAR DIS ENT IDM SKEKUR STD 696 694 10033 505 237 9521 9769 9643 9311
8 MNI VAR DIS ENT IND CORUNI SKE STD 695 699 10029 451 295 9570 9714 9641 9308
9 MNI VAR DIS ENT IND CORIDM SKE STD 688 690 10035 457 283 9564 9726 9644 9313
10 MNI VAR DIS ENT COR IDMSKE KUR STD 689 691 10036 478 261 9545 9747 9645 9314
PrecisionSensitivity
F-measureAccuracy
90919293949596979899
Perfo
rman
ce (
)
2 3 4 5 6 7 8 9 101Feature combination index
Figure 6 Results of performance at each feature combination in thebehavioral solution set
to distinguish The best solution was determined to be indexnumber 2 combination of MNI DIS COR SKE and STDwhich has the lowest of resubstitution and cross-validationerrors 668 and 670 respectively This combinationincludes all of the top features based on the probability occur-rence except ENT and VAR The four performance results ateach feature combination in the behavioral solution set areshown in Figure 6 where the highest results aremarked in redcircles and the lowest in green-dot circles Sensitivity appearshigher than precision at each feature combination becauseFPs are larger than FNs in the confusion matrix Particularlyindex number 7 has the biggest difference between FP and FNresulting in the highest sensitivity and lowest precision Thesummation of FP and FN in index number 4 is the highestin the behavioral solution set resulting in the lowest accuracywhile index number 2 has the lowest summation of FP andFN providing the highest accuracy
This study investigated a wide range of features fromlong-wavelength infrared camera images analyzed normaldistributions of fifteen features with respect to the classes ofsmoke fire and their thermal reflections and discovered thehighest performing feature combination by examining singlefeatures and multiple feature combinations As a result theproposed feature combination of MNI DIS COR SKE andSTD increases the performance compared with the previousstudy [1] which used MNI VAR ENT and IDM As shownin Figure 7 the errors are reduced by 286 and 268 resub-stitution and cross-validation errors and performances areincreased by 290 158 020 and 285 accuracy F-measure sensitivity and precision respectively
Figure 8 shows original visual and thermal images withthe robot at three different locations start point hallwayentrance and room entrance described in the experimentalfacility Each row relates to a series of images from the robotat three locations The first row contains visible images ofthe robot view As seen in the visible image at start pointfurther information regarding the hallway is limited due toshadowing of the light The image at hallway entrance showsa smoke layer in the upper portion of the hallway due to afire inside the roomThe image at the room entrance displaysa wood crib fire with sparks Because of soot and relativedifference in brightness the background is shown darker andthus limiting information on the background around the fire
Thermal infrared images are displayed in the second rowto show information that RGB camera cannot provide infire environments Unlike visual image at start point thatis obscured due to shadowing the presence of smoke andits thermal reflections on the ventilation hood can be obvi-ously perceived The red boxes on thermal images indicateobjects extracted through the adaptive object extraction withoptical flows and identification numbers In spite of dense
Journal of Sensors 9
9279
9741
9504
9055
954
938
9564
9761
9662
9345
668
670
Precision
Sensitivity
F-measure
Accuracy
Resubstitution error
Cross-validation error
The proposedThe previous
2000 4000 6000 8000 10000000()
Figure 7 Result comparison between the previous and the proposed studies
① Start point ② Hallway entrance ③ Room entrance
IR im
age
class
ifica
tion
resu
lts
Visib
le im
age
IR im
age
imag
e ana
lysis
Figure 8 Original visible and IR images with image analysis and classification results at different locations in the test setup of actual fires
smoke-filled and low visibility environments thermal imagescan generate the images of smoke and fire as well as back-ground information that is otherwise not visible throughvisual imaging
On the third row class labels and posterior probabilitiesof each candidate are displayed at the center of candidateROI as a result of Bayesian classification Using enhancedimage processing techniques the thermal images can bemore
10 Journal of Sensors
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
2
0
19
4820
64
Feature combination index number 2
Feature combination index number 3 Feature combination index number 4
Feature combination index number 5
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1462
0
0
0
1
1462
0
0
0
1
1457
0
0
0
1
1457
0
0
1365
193
197
0
0
1343
181
196
0
0
1325
205
181
0
0
1270
207
181
0
0
53
167
1929
0
0
145
166
1943
0
0
101
165
1942
7
0
49
4844
62
2
0
30
4817
63
0
0
53
165
1929
7
0
27
4830
61
MNI DIS COR SKE STD
MNI ENT COR SKE STD MNI VAR DIS ENT CON COR SKE STD
MNI VAR DIS ENT COR UNI SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bjPredicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Feature combination index number 1
140
159
0
0
0
0
0
00
0
0
0
0
0
1981
488
1280
193
144
1
1462
4838
62
25
2
MNI VAR ENT COR SKE
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Feature combination index number 6
0
0
0
0
488
0
0
0
1
14620
0
1339
206
185
0
0
93
161
1935
2
0
13
4823
67
MNI VAR DIS ENT COR IDM SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
F-re
flect
ion
Oth
er o
bj
Smok
e
Fire
S-re
flect
ion
Predicted class
Figure 9 Continued
Journal of Sensors 11
Feature combination index number 7 Feature combination index number 8
Feature combination index number 9 Feature combination index number 10
89
133
0
0
0
0
0
02
0
0
0
0
0
1888
488
0
0
0
0
488
1341
200
235
1
1459
4857
62
15
5
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1458
0
0
0
1
1462
0
0
0
1
1462
0
0
1324
204
181
0
0
1334
208
181
0
0
1334
205
204
0
0
99
174
1943
0
0
91
150
1921
2
0
18
4810
65
6
0
20
4835
62
0
0
93
172
1941
2
0
22
4812
63
MNI VAR DIS ENT COR IDM SKE KUR STD
MNI VAR DIS ENT IDM SKE KUR STD
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bj
Predicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
MNI VAR DIS ENT IND COR IDM SKE STD
MNI VAR DIS ENT IND COR UNI SKE STD
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
Figure 9 Results of confusion matrix at each feature combination
refined and clearer than the thermal images on the secondrow Smoke fire and their thermal reflections are identifiedand marked in red or orange ellipses
6 Conclusion
The appropriate combination of features was investigated toaccurately classify fire smoke and their thermal reflectionsusing thermal images Gray-scale 14-bit images from a singleinfrared camera were used to extract motion and texture fea-tures by applying a clustering-based autothresholding tech-nique Bayesian classification is performed to probabilisti-cally identify multiple classes during real-time implemen-tation To find the best combination of features a multi-objective genetic algorithm optimization was implementedusing resubstitution and cross-validation errors as objectivefunctions Large-scale fire tests with different fire sources
were conducted to create a range of temperature and smokeconditions to evaluate the feature combinations
Fifteenmotion and texture features were analyzed and theprobability density functions of the features were computedby the maximum likelihood estimation The combinationof multiple features was determined to more accuratelyclassify fire smoke and thermal reflections compared witha single feature In the behavioral solution set where featurecombinations produce less than 7 resubstitution and cross-validation errors COR ENT DIS SKE STD VAR andMNIhad 800 or more occurrence while other features had400 or less occurrence The feature combination of MNIDIS COR SKE and STD produced the highest performancein the classification resulting in 668 and 670 resubstitu-tion and cross-validation errors and 9564 9761 9662and 9345 precision sensitivity F-measure and accuracyrespectively
12 Journal of Sensors
In the near future the classification of fire smoke andtheir thermal reflections will be evaluated on any classifiersand features to increase performanceThe convolution neuralnetwork of deep learning which has recently shown highperformance could be explored as a classifier also model-based image features such as discrete wavelet transform willbe further studied
Appendix
See Figure 9
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this manuscript
Acknowledgments
This work was sponsored by the Office of Naval ResearchGrant no N00014-11-1-0074 scientific office Dr ThomasMcKenna in USA Hwarang-dae Research Institute in Seouland Agency for Defense Development in Daejeon SouthKorea The authors would like to thank Mr Joseph Starr andMr JoshMcNeil for assisting in performing the fire testsTheauthors would also like to thank Rosana K Lee for helpingand supporting this research
References
[1] J-H Kim and B Y Lattimer ldquoReal-time probabilistic classifi-cation of fire and smoke using thermal imagery for intelligentfirefighting robotrdquo Fire Safety Journal vol 72 pp 40ndash49 2015
[2] J W Starr and B Y Lattimer ldquoEvaluation of navigation sensorsin fire smoke environmentsrdquo Fire Technology vol 50 no 6 pp1459ndash1481 2014
[3] J-H Kim B Keller and B Y Lattimer ldquoSensor fusion basedseek-and-find fire algorithm for intelligent firefighting robotrdquoin Proceedings of the IEEEASME International Conference onAdvanced Intelligent Mechatronics (AIM rsquo13) pp 1482ndash1486IEEE Wollongong Australia July 2013
[4] J G McNeil J Starr and B Y Lattimer ldquoAutonomous fire sup-pression using multispectral sensorsrdquo in Proceedings of theIEEEASME International Conference on Advanced IntelligentMechatronics Mechatronics for HumanWellbeing (AIM rsquo13) pp1504ndash1509 Wollongong Austrailia July 2013
[5] J-H Kim J W Starr and B Y Lattimer ldquoFirefighting robotstereo infrared vision and radar sensor fusion for imagingthrough smokerdquo Fire Technology vol 51 no 4 pp 823ndash8452015
[6] RC Luo andK L Su ldquoAutonomous fire-detection systemusingadaptive sensory fusion for intelligent security robotrdquo IEEEASME Transactions on Mechatronics vol 12 no 3 pp 274ndash2812007
[7] M A Jackson and I Robins ldquoGas sensing for fire detectionmeasurements of CO CO
2
H2
O2
and smoke density in Euro-pean standard fire testsrdquo Fire Safety Journal vol 22 no 2 pp181ndash205 1994
[8] B U Toreyin R G Cinbis Y Dedeoglu and A E Cetin ldquoFiredetection in infrared video using wavelet analysisrdquo OpticalEngineering vol 46 no 6 Article ID 067204 2007
[9] B U Toreyin Y Dedeoglu andA E Cetin ldquoWavelet based real-time smoke detection in videordquo in Proceedings of the 13th Euro-pean Signal Processing Conference pp 4ndash8 Antalya TurkeySeptember 2005
[10] T Celik H Demirel H Ozkaramanli and M Uyguroglu ldquoFiredetection using statistical color model in video sequencesrdquoJournal of Visual Communication and Image Representation vol18 no 2 pp 176ndash185 2007
[11] L Merino F Caballero J R Martınez-de-Dios I Maza and AOllero ldquoAn unmanned aircraft system for automatic forest firemonitoring and measurementrdquo Journal of Intelligent amp RoboticSystems vol 65 no 1 pp 533ndash548 2012
[12] Y Wang T W Chua R Chang and N T Pham ldquoReal-timesmoke detection using texture and color featuresrdquo in Proceed-ings of the 21st International Conference on Pattern Recognition(ICPR rsquo12) pp 1727ndash1730 Tsukuba Japan November 2012
[13] G Marbach M Loepfe and T Brupbacher ldquoAn image process-ing technique for fire detection in video imagesrdquo Fire SafetyJournal vol 41 no 4 pp 285ndash289 2006
[14] M I Chacon-Murguia and F J Perez-Vargas ldquoThermal videoanalysis for fire detection using shape regularity and intensitysaturation featuresrdquo in Pattern Recognition J F Martınez-Trinidad J A Carrasco-Ochoa C B-Y Brants and E RHancock Eds vol 6718 of Lecture Notes in Computer Sciencepp 118ndash126 Springer Berlin Germany 2011
[15] W Phillips III M Shah and N D V Lobo ldquoFlame recognitionin videordquo Pattern Recognition Letters vol 23 no 1ndash3 pp 319ndash327 2002
[16] D Han and B Lee ldquoDevelopment of early tunnel fire detectionalgorithm using the image processingrdquo in Advances in VisualComputing pp 39ndash48 Springer Berlin Germany 2006
[17] Y Chunyu F Jun W Jinjun and Z Yongming ldquoVideo firesmoke detection using motion and color featuresrdquo Fire Technol-ogy vol 46 no 3 pp 651ndash663 2010
[18] F Yuan ldquoVideo-based smoke detection with histogram se-quence of LBP and LBPV pyramidsrdquo Fire Safety Journal vol 46no 3 pp 132ndash139 2011
[19] F Lafarge X Descombes and J Zerubia ldquoTextural kernel forSVM classification in remote sensing application to forest firedetection andUrban area extractionrdquo in Proceedings of the IEEEInternational Conference on Image Processing (ICIP rsquo05) pp1096ndash1099 September 2005
[20] F Amon and A Ducharme ldquoImage frequency analysis for test-ing of fire service thermal imaging camerasrdquo Fire Technologyvol 45 no 3 pp 313ndash322 2009
[21] F Amon V Benetis J Kim and A Hamins ldquoDevelopmentof a performance evaluation facility for fire fighting thermalimagersrdquo in Defense and Security pp 244ndash252 2004
[22] F D Maxwell ldquoA portable IR system for observing fire thrusmokerdquo Fire Technology vol 7 no 4 pp 321ndash331 1971
[23] A Barducci D Guzzi P Marcoionni and I Pippi ldquoInfrareddetection of active fires and burnt areas theory and observa-tionsrdquo Infrared Physics amp Technology vol 43 no 3ndash5 pp 119ndash125 2002
[24] C Wang and S Qin ldquoAdaptive detection method of infraredsmall target based on target-background separation via robustprincipal component analysisrdquo Infrared Physics amp Technologyvol 69 pp 123ndash135 2015
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
Journal of Sensors 3
(a) RGB images
(b) Thermal images
Figure 2 (a) RGB images of fire scenes and (b) extracted objectsfrom thermal images with optical flow vectors overlaid
motion and texture features were selected because theywere crucial in the previous study of fire andor smokedetection and also best suitable for the thermal image analysisthat is major information the firefighting robot can acquireunder fire environments Optical flow a popular motionmeasurement was used for the motion features while thefirst- and second-statistical texture features were applied forthe texture measurement
A FLIR A35 long-wavelength IR camera which is capableof imaging through zero visibility environments was usedto produce images All images were from a 320 times 256-pixelfocal plane array 60Hz frame rate that produces 14-bit imageswith an intensity range of minus16384 for minus40∘C to minus1 for 550∘CFifteen features from optical flow and the statistical texturefeatures are evaluated to find the best feature combinationOptical flow shows temporal variations due tomoving objectsin the FOV or motion of the robot The first- and second-order statistical texture features display spatial characteristicsof objects in the scene
31 Motion Features by Optical Flow Optical flow is a usefultool to recognize motion of an object in sequential images[30] It consists of local and global methods Lucas-Kanade(LK) is a local method that is relatively robust with a lessdense flowfield whileHorn-Schunck (HS) is a globalmethodwith a dense flow field and high sensitivity to noise [31]Because the intensities in the thermal image change due tothe varying fire environment LK method that has higherrobustness compared with HS was selected in this researchto measure motion features of the objects Two featuresof optical flow vector number (OFVN) and optical flowmeanmagnitude (OFVMM)were computed to quantitativelycharacterize motions of fire smoke and their reflectionsFigure 2 contains RGB and thermal images of dense smokein a hallway and a wood crib fire in a room Red arrows in the
thermal images indicate the direction and magnitude of theoptical flow vectors with red boxes that show smoke fire andthermal reflections
32 First- and Second-Order Statistical Texture Features Thefirst- and second-order statistical features were considered inthis study for object classification The first-order statisticalfeatures estimate individual property of pixels not character-izing any relationship between neighboring pixels and canbe computed using the intensity histogram of the candidateregion of interest (ROI) in the image As described in [32]mean (MNI) variance (VAR) standard deviation (STD)skewness (SKE) and kurtosis (KUR) were calculated by
MNI = 1
119873119875
119873119875
sum
119894119895=1
119868119894119895
VAR = 1
119873119875
119873119875
sum
119894119895=1
(119868119894119895minus 120583)
2
STD (120590) = ( 1
119873119875
119873119875
sum
119894119895=1
(119868119894119895minus 120583)
2
)
12
SKE = 1
1205903
119873119875
119873119875
sum
119894119895=1
(119868119894119895minus 120583)
3
KUR = 1
1205904
119873119875
119873119875
sum
119894119895=1
(119868119894119895minus 120583)
4
(1)
where 119868119894119895
refers to the intensity of a pixel at 119894 and 119895 and119873119875denotes the number of pixels (NOP) of the object in the
image The second-order statistical features represent spatialrelationships between a pixel and its neighbors Gray-levelcooccurrence matrix (GLCM) [33] is used to account foradjacent pixel relationships in four directions (horizontalvertical left and right diagonals) by quantizing the spatialcooccurrence of neighboring pixels A total of seven second-order statistics features were used including dissimilarity(DIS) entropy (ENT) contrast (CON) inverse difference(INV) correlation (COR) uniformity (UNI) and inverse dif-ferencemoment (IDM) Tomeasure these features a normal-ized cooccurrence matrix 119862
119894119895is used which can be defined as
119862119894119895=
119875119894119895
sum119873119866
119894119895=1
119875119894119895
(2)
where 119875119894119895refers to the frequency of occurrences of the gray-
level of adjacent pixels at 119894 and 119895 within the four directionsand 119873
119866denotes the number of the gray-levels in the
quantized image The denominator of (2) normalizes 119875119894119895to
be estimates of the cooccurrence probabilities After buildingthe normalized cooccurrence matrix 119862
119894119895 seven features of
the second-order statistics features were computed by
DIS = sum119862119894119895
1003816100381610038161003816119894 minus 119895
1003816100381610038161003816
ENT = minussum119862119894119895log119862119894119895
4 Journal of Sensors
CON = sum119862119894119895(119894 minus 119895)
2
INV = sum119862119894119895
1 +1003816100381610038161003816119894 minus 119895
1003816100381610038161003816
COR = sum(119894119895 + 120583
119894120583119895minus 120583119894minus 120583119895) 119862119894119895
120590119894120590119895
UNI = sum1198622119894119895
IDM = sum
119862119894119895
1 + (119894 minus 119895)2
(3)
4 Object Extraction andBayesian Classification
One of the main characteristics of fire smoke and theirthermal reflections in thermal images is that they are higherin intensity than the background With intensity related totemperature in the thermal image higher temperature objectsappear brighter than the background Hence intensity is aprimary factor for object extraction from the backgroundAssuming that the thermal image histogram has a bimodaldistribution for foreground (ie object) and backgroundthe clustering-based image autothresholding method [34]called Otsumethod can calculate an optimum threshold thatseparates objects and background creating a binary imagewith 0 being the background and 1 being the objects Thebinary images were filtered to remove small regions and holesinside objects through morphological filtering techniquesAfter convoluting the original 14-bit image with the filtered-binary image a final image was obtained that includes theoriginal 14-bit intensities in objects as well as zeroes in thebackground
There are several classification methods commonly usedin supervised machine learning 119896-nearest neighbors (119896NN)decision tree (DT) neural networks (NN) support vectormachine (SVM) and Naıve Bayesian For this study theseclassification methods were analyzed by considering threepoints capability to classify multiple classes such as firesmoke and their thermal reflections less chance of overfit-ting problem because under fire environments there couldbe a number of situations that are not learned or trainedreal-time implementation because firefighting robot needs tomake a decision in real-time otherwise it cannot operate itstask 119896NN is insensitive to outliers but it needs a large amountof memory and expensive computation [35] DT has lowcomputation burden but for the multiclasses classificationit may generate a complicated tree structure and may causeoverfitting problem [35 36] NN shows high performancewhen processing with multidimensions and continuous fea-tures but cannot overcome overfitting problem SVM pro-vides fast computation and the highest accuracy but it cannotbe used for the multilabel classification because it producesbinary results [37] Naıve Bayesian classification is Bayesrsquotheorem-based probabilistic classification and is popular for
pattern recognition applications Although this method haslower accuracy compared with other classifiers and assumesthat each feature is independent it has fast computationrobustness to untrained cases and less chance of overfitting[35] In addition this classification has the capability ofprobabilistic decision making over multiple classes with fastcomputation for real-time implementation In this studyBayesian classification is used for evaluation of each feature
With several given features 1198651 1198652 119865
119902 (motion and
texture features) we can calculate the probability that oneclass 119862
ℎ(fire smoke thermal reflections etc) corresponds
to the candidate 119896 by using a conditional probability 119896119901(119862ℎ|
11986511198652sdot sdot sdot 119865119902) also known as the posterior probability By using
Bayesrsquo theorem it can be written with prior likelihood andevidence as shown in
119896
119901 (119862ℎ| 11986511198652sdot sdot sdot 119865119902)
=
119896
119901 (119862ℎ)119896
119891 (11986511198652sdot sdot sdot 119865119902| 119862ℎ)
sum119862ℎ
119896119901 (119862ℎ)119896
119891 (11986511198652sdot sdot sdot 119865119902| 119862ℎ)
(4)
where 119896119901(119862ℎ) is the prior probability meaning it represents
candidate 119896 probability to be 119862ℎand can be calculated by
number of samples of class 119862ℎdivided by the total number
of samples 119896119891(11986511198652sdot sdot sdot 119865119902| 119862ℎ) is the likelihood function
and the denominator of (4) is the evidence that plays asa normalizing constant by the summation of productionbetween the prior and likelihood at each class By applying theconditional independence assumption the likelihood func-tion can be rewritten by
119891 (11986511198652sdot sdot sdot 119865119902| 119862ℎ) =
119902
prod
119894=1
119891 (119865119894| 119862ℎ) (5)
The conditional probability density function 119891(119865119894| 119862ℎ) sim
119873(120583119865119894|119862ℎ
Σ119865119894|119862ℎ
) can be described as
119891 (119865119894| 119862ℎ) =
1
radic2120587Σ119865119894|119862ℎ
119890minus(12)(119865119894minus120583119865119894|119862ℎ
)
119879Σ
minus1
119865119894|119862ℎ(119865119894minus120583119865119894|119862ℎ
)
(6)
where
120583119865119894|119862ℎ
= 120583119865119894+ Σ119879
119862ℎ119865119894
Σminus1
119862ℎ119862ℎ
(119862ℎminus 120583119862ℎ)
Σ119865119894|119862ℎ
= Σ119865119894119865119894minus Σ119879
119862ℎ119865119894
Σminus1
119862ℎ119862ℎ
Σ119862ℎ119865119894
(7)
As shown in Table 2 Gaussian parameters for fifteen featureswith respect to smoke smoke thermal reflection fire andfire thermal reflection were estimated by using the maximumlikelihood estimation [38] Probability density distributionsfor the entire features are illustrated in Figure 3 With (5) theevidence and then the posterior probability of each class werecalculated By applying the maximum priority decision rulein (8) the Bayesian classification was used to predict the classand probability of each candidate in the scene
class = argmax119862ℎ
(119901 (119862ℎ| 11986511198652sdot sdot sdot 119865119902))
prob = max (119901 (119862ℎ| 11986511198652sdot sdot sdot 119865119902))
(8)
Journal of Sensors 5
Table 2 Gaussian parameters
Smoke Smoke-reflection Fire Fire-reflection120583 Σ 120583 Σ 120583 Σ 120583 Σ
MNI minus12665119864 + 04 69008119864 + 02 minus13383119864 + 04 35543119864 + 02 minus59714119864 + 03 16050119864 + 03 minus71399119864 + 03 46070119864 + 02
VAR 47578119864 + 05 44154119864 + 05 44012119864 + 04 41227119864 + 04 10501119864 + 07 22343119864 + 06 13300119864 + 06 88178119864 + 05
NOP 10733119864 + 03 42287119864 + 03 32220119864 + 01 13676119864 + 02 22174119864 + 02 17105119864 + 03 59575119864 + 01 23911119864 + 02
STD 62146119864 + 02 29931119864 + 02 18314119864 + 02 10236119864 + 02 32170119864 + 03 38930119864 + 02 10534119864 + 03 46992119864 + 02
SKE 11672119864 minus 01 58208119864 minus 01 minus94009119864 minus 02 62857119864 minus 01 22230119864 minus 02 52287119864 minus 01 22385119864 minus 01 71380119864 minus 01
KUR 30045119864 + 00 20213119864 + 00 36553119864 + 00 16131119864 + 00 22283119864 + 00 11517119864 + 00 34848119864 + 00 11912119864 + 00
OFVN 28832119864 + 04 14156119864 + 04 11848119864 + 03 87345119864 + 02 11257119864 + 04 13722119864 + 04 29978119864 + 03 22034119864 + 03
OFVMM 86830119864 + 01 56259119864 + 01 11687119864 + 02 79444119864 + 01 17598119864 + 02 29308119864 + 02 12641119864 + 02 67234119864 + 01
DIS 71791119864 minus 02 17966119864 minus 02 28045119864 minus 02 12245119864 minus 02 10937119864 minus 01 86515119864 minus 02 47308119864 minus 02 28550119864 minus 02
ENT 41949119864 minus 01 93683119864 minus 02 19576119864 minus 01 86106119864 minus 02 33681119864 minus 01 19107119864 minus 01 17115119864 minus 01 99702119864 minus 02
CON 89384119864 minus 01 36871119864 minus 01 12402119864 minus 01 65921119864 minus 02 92956119864 minus 01 69179119864 minus 01 39683119864 minus 01 26996119864 minus 01
IND 98588119864 minus 01 63845119864 minus 03 99646119864 minus 01 15141119864 minus 03 96334119864 minus 01 37231119864 minus 02 98771119864 minus 01 79275119864 minus 03
COR 96636119864 minus 01 31625119864 minus 02 88533119864 minus 01 43467119864 minus 02 89406119864 minus 01 30935119864 minus 02 88417119864 minus 01 57861119864 minus 02
UNI 51044119864 minus 01 20109119864 minus 01 95222119864 minus 01 31669119864 minus 02 67906119864 minus 01 26718119864 minus 01 85305119864 minus 01 10599119864 minus 01
IDM 60061119864 minus 01 19160119864 minus 01 97533119864 minus 01 16844119864 minus 02 77850119864 minus 01 24321119864 minus 01 92043119864 minus 01 58631119864 minus 02
Table 3 The object numbers of smoke smoke-reflection fire fire-reflection and other hot objects classes
Type Total Smoke Smoke-reflection Fire Fire-reflection Other hot objectsNumber of objects 10775 5190 1445 1464 489 2187
Figure 3 shows probability density distribution of eachclass using the Gaussian parameters of Table 2 Gaussiandistribution for classes in Figure 3 shows how fire fire-reflection smoke and smoke-reflection are distributed by thefifteen features Some features split out the distribution of thefour classes while others cause overlap For example MNIbest describes a well split out case of the classes althoughsmoke and its reflection and fire and its reflection do overlapSKE shows the worst case in which all classes overlap makingit impossible to distinguish any of the four classes
5 Result and Discussion
The accuracy in classifying fire objects was analyzed usingdata from a series of large-scale tests in the facility [1]using actual fires up to 75 kW Fires included latex foamwood cribs and propane gas fires from a sand burner Thesedifferent types of fires produced a range of temperature andsmoke conditions Latex foam fires produced lower tempera-ture conditions but dense low visibility smoke Converselypropane gas fires produced higher gas temperatures andlight smoke Wood crib fires resulted in smoke and gastemperatures between those of latex foam and propane gasfires however these fires resulted in sparks created from theburning wood Thermal images were collected by driving awheeled mobile robot through the setup during a fire test Atotal of 10775 objects were collected from the experimentsand categorized as either smoke smoke-reflection fire fire-reflection or other hot objects in order to be served as cluesto lead the firefighting robot to navigate toward the fire source
outside the FOV In addition as each object has sixteen cor-responding data points (fifteen features and a class) the totalnumber of data points used in this paper is 172400The num-bers of each object in this experiment are shown in Table 3
Two types of error criterions (resubstitution and 119896-foldcross-validation errors [39]) were used to measure howeach feature accurately performs in the classification Resub-stitution error takes the entire dataset to compare the actualclasses with the predicted classes by the Bayesian classifica-tion in order to examine how well the actual and predictedclasses match each other When this criterion is used aloneto enhance accuracy the classification can be overfitted tothe training dataset Cross-validation error is advantageousto detect and prevent from overfitting Instead of using theentire dataset cross-validation randomly selects and splitsthe dataset into 119896 partitions of approximately equal size(119896 = 10) to estimate a mean error by comparing betweenthe randomly selected partition and trained results of theremaining partitions
51 Single Feature Performance The performance results ofeach feature are shown in Table 4 The first-order statisticaltexture features MNI VAR and STD produced the lowesterrors while NOP SKE and KUR show the highest Theseresults show that MNI and VAR are beneficial to distinguishfire smoke and thermal reflections while motion features arenot As NOP shows the highest error OFVMM one of themotion features shows the second highest errors comparedwith the other features This is in part attributed to thedynamic motion of the robot ENT and COR second-order
6 Journal of Sensors
0 5 10VAR
Prob
abili
ty d
ensit
y
SmokeS-reflection
FireF-reflection
times10minus4
times10minus4
times104
times10minus7
times106
times10minus3
2468
1012
Prob
abili
ty d
ensit
y
minus10000 minus8000 minus6000 minus4000minus12000MNI
0
5
10
Prob
abili
ty d
ensit
y
0
05
1
Prob
abili
ty d
ensit
y
5000 100000OFVN
0
20
40
Prob
abili
ty d
ensit
y
01 02 03 040DIS
500 1000 1500 20000OFVMM
0
0005
001
Prob
abili
ty d
ensit
y0
5
Prob
abili
ty d
ensit
y
2 4 60NOP
50100150200250
Prob
abili
ty d
ensit
y
09 095 1085IND
1 15 2 2505CON
2
4
6
Prob
abili
ty d
ensit
y
02 04 06 08 10ENT
0
5
Prob
abili
ty d
ensit
y
2468
1012
Prob
abili
ty d
ensit
y
06 08 104COR
minus2 0 2 4 6minus4SKE
0
05
1
Prob
abili
ty d
ensit
y
01
02
03
Prob
abili
ty d
ensit
y
5 10 15 200KUT
1000 2000 3000 40000STD
0
2
4
08 10604IDM
510152025
Prob
abili
ty d
ensit
y
02 04 06 08 10UNI
0
10
20
Prob
abili
ty d
ensit
y
SmokeS-reflection
FireF-reflection
SmokeS-reflection
FireF-reflection
times10minus3
Figure 3 Probability density distributions of each feature
statistical texture features show42sim45error which is higherthan the other second-order features
52 Multiple Feature Combination Performance The errorresults in Table 4 demonstrate that a single feature cannotaccurately classify fire smoke and thermal reflections Thuspossible combination ofmultiple features was considered andanalyzed to find the best combination of the featuresThe totalnumber of all possible combinations that have two or morefeatures is
119873total =119898
sum
119911=2
(
119898
119911
) (9)
where 119898 refers to the total number of features (ie 119898 = 15)and 119911 is the number of features in the combination Basedon all possible combination the multiobjective genetic algo-rithm optimization [40] in the global optimization toolbox ofMATLAB was used to find the best combination of featuresthat has the highest performance in the classification Theobjective functions in the optimization resubstitution and119896-fold cross-validation errors [39] were used to measurehow accurately different feature combinations perform in theclassification
Figure 4 contains a plot of the error associated with themost promising feature combinations The behavioral solu-tion set is defined as feature combinations with less than 7
Journal of Sensors 7
Table 4 Performance of each feature
Resubstitution error () Cross-validation error ()MNI 237 238VAR 243 243NOP 721 720STD 232 232SKE 527 528KUR 506 506OFVN 395 400OFVMM 586 586DIS 290 290ENT 446 446CON 285 285IND 301 301COR 412 411UNI 345 345IDM 377 377
Behavioral region
Behavioral solution setGeneral set
66 67 68 69 7 71 72 7365Resubstitute error ()
65
66
67
68
69
7
71
72
73
Cros
s-va
lidat
ion
erro
r (
)
Figure 4Multiobjective optimization result showing the general setand behavioral solution set (colored region)
error for both objective functions while the general set refersto all other possible feature combinations The behavioralsolution set contains 00061 of all possible feature combi-nations
The occurrence probability of features in the behavioralsolution set is illustrated in Figure 5 In the behavioralsolution set the first-order statistic texture features MNI andSKE always exist while OFVN NOP and OFVMM featuresdo not Both the first-order statistical texture features STDand VAR and the second-order statistical texture featuresCOR ENT andDIS showahigher occurrence comparedwiththe other first- and second-order texture features while KURIDMUNI IND andCON show lower occurrence Note thatdue to the robotrsquos dynamical motion motion features werenot successful and even not included in the top 10 featurecombinations of the behavioral set
1000800
00900
1000200
0000
800900
100200
900200
400
MNIVARNOPSTDSKE
KUROFVN
OFMMDIS
ENTCONINDCORUNIIDM
200 400 600 800 1000 120000Occurrence probability ()
Figure 5 Occurrence analysis of the features in the behavioralregion
The top features based on the probability occurrence inFigure 5 are COR ENT DIS SKE STD VAR and MNIHowever the combination of these seven features does notresult in the best solution for classification Table 5 containsthe classification performance of the combination of featuresin the behavioral solution set In order to evaluate the per-formance of each feature combination various performancemeasures have been used such as precision sensitivity F-measure and accuracy Precision measures the fraction ofpositive instances from the group that the classifier predictedto be positive and recall measures the fraction of positiveexamples from the positive group of the actual class and [35]F-measure is the harmonic mean and accuracy is the pro-portion of true resultsThesemeasures can bemathematicallydefined as
Precision = TPTP + FP
Sensitivity = TPTP + FN
119865-Measure =2 (Sensitivity sdot Precision)Sensitivity + Precision
Accuracy = TPTP + FP + FN
(10)
where TP is correctly classified positive cases FP is incor-rectly classified negative cases and FN is incorrectly classifiedpositive cases For the performance measurement confusionmatrixes were created as described in Appendix and appliedinto (10) In the precision index number 1 combinationshows the highest performance in the behavioral solutionset while index number 7 combination shows the lowestIn the sensitivity index number 7 combination records thehighest results while index number 4 does the lowest In theF-measure and accuracy index number 2 combination showsthe highest record while index number 4 does the lowestBased on the confusion matrixes most of misclassificationoccurs in the classification of smoke smoke-reflection andother hot objects because during small fire texture patternsof these classes were diminished and the intensity was too low
8 Journal of Sensors
Table 5 Results of error case and performance at each feature combination in the behavioral solution set (Resu means resubstitution andCros refers to cross-validation)
Index Combination of features Error () Case Performance ()Resu Cros TP FP FN Precision Sensitivity 119865-measure Accuracy
1 MNI VAR ENT COR SKE 690 689 10049 402 324 9615 9688 9651 93262 MNI DIS COR SKE STD 668 670 10069 459 247 9564 9761 9662 93453 MNI ENT COR SKE STD 676 675 10061 447 267 9575 9741 9657 9337
4 MNI VAR DIS ENT CON CORSKE STD 698 690 9980 454 341 9565 9670 9617 9262
5 MNI VAR DIS ENT COR UNISKE STD 689 690 10037 453 285 9568 9724 9645 9315
6 MNI VAR DIS ENT COR IDMSKE STD 677 679 10047 461 267 9561 9741 9650 9324
7 MNI VAR DIS ENT IDM SKEKUR STD 696 694 10033 505 237 9521 9769 9643 9311
8 MNI VAR DIS ENT IND CORUNI SKE STD 695 699 10029 451 295 9570 9714 9641 9308
9 MNI VAR DIS ENT IND CORIDM SKE STD 688 690 10035 457 283 9564 9726 9644 9313
10 MNI VAR DIS ENT COR IDMSKE KUR STD 689 691 10036 478 261 9545 9747 9645 9314
PrecisionSensitivity
F-measureAccuracy
90919293949596979899
Perfo
rman
ce (
)
2 3 4 5 6 7 8 9 101Feature combination index
Figure 6 Results of performance at each feature combination in thebehavioral solution set
to distinguish The best solution was determined to be indexnumber 2 combination of MNI DIS COR SKE and STDwhich has the lowest of resubstitution and cross-validationerrors 668 and 670 respectively This combinationincludes all of the top features based on the probability occur-rence except ENT and VAR The four performance results ateach feature combination in the behavioral solution set areshown in Figure 6 where the highest results aremarked in redcircles and the lowest in green-dot circles Sensitivity appearshigher than precision at each feature combination becauseFPs are larger than FNs in the confusion matrix Particularlyindex number 7 has the biggest difference between FP and FNresulting in the highest sensitivity and lowest precision Thesummation of FP and FN in index number 4 is the highestin the behavioral solution set resulting in the lowest accuracywhile index number 2 has the lowest summation of FP andFN providing the highest accuracy
This study investigated a wide range of features fromlong-wavelength infrared camera images analyzed normaldistributions of fifteen features with respect to the classes ofsmoke fire and their thermal reflections and discovered thehighest performing feature combination by examining singlefeatures and multiple feature combinations As a result theproposed feature combination of MNI DIS COR SKE andSTD increases the performance compared with the previousstudy [1] which used MNI VAR ENT and IDM As shownin Figure 7 the errors are reduced by 286 and 268 resub-stitution and cross-validation errors and performances areincreased by 290 158 020 and 285 accuracy F-measure sensitivity and precision respectively
Figure 8 shows original visual and thermal images withthe robot at three different locations start point hallwayentrance and room entrance described in the experimentalfacility Each row relates to a series of images from the robotat three locations The first row contains visible images ofthe robot view As seen in the visible image at start pointfurther information regarding the hallway is limited due toshadowing of the light The image at hallway entrance showsa smoke layer in the upper portion of the hallway due to afire inside the roomThe image at the room entrance displaysa wood crib fire with sparks Because of soot and relativedifference in brightness the background is shown darker andthus limiting information on the background around the fire
Thermal infrared images are displayed in the second rowto show information that RGB camera cannot provide infire environments Unlike visual image at start point thatis obscured due to shadowing the presence of smoke andits thermal reflections on the ventilation hood can be obvi-ously perceived The red boxes on thermal images indicateobjects extracted through the adaptive object extraction withoptical flows and identification numbers In spite of dense
Journal of Sensors 9
9279
9741
9504
9055
954
938
9564
9761
9662
9345
668
670
Precision
Sensitivity
F-measure
Accuracy
Resubstitution error
Cross-validation error
The proposedThe previous
2000 4000 6000 8000 10000000()
Figure 7 Result comparison between the previous and the proposed studies
① Start point ② Hallway entrance ③ Room entrance
IR im
age
class
ifica
tion
resu
lts
Visib
le im
age
IR im
age
imag
e ana
lysis
Figure 8 Original visible and IR images with image analysis and classification results at different locations in the test setup of actual fires
smoke-filled and low visibility environments thermal imagescan generate the images of smoke and fire as well as back-ground information that is otherwise not visible throughvisual imaging
On the third row class labels and posterior probabilitiesof each candidate are displayed at the center of candidateROI as a result of Bayesian classification Using enhancedimage processing techniques the thermal images can bemore
10 Journal of Sensors
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
2
0
19
4820
64
Feature combination index number 2
Feature combination index number 3 Feature combination index number 4
Feature combination index number 5
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1462
0
0
0
1
1462
0
0
0
1
1457
0
0
0
1
1457
0
0
1365
193
197
0
0
1343
181
196
0
0
1325
205
181
0
0
1270
207
181
0
0
53
167
1929
0
0
145
166
1943
0
0
101
165
1942
7
0
49
4844
62
2
0
30
4817
63
0
0
53
165
1929
7
0
27
4830
61
MNI DIS COR SKE STD
MNI ENT COR SKE STD MNI VAR DIS ENT CON COR SKE STD
MNI VAR DIS ENT COR UNI SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bjPredicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Feature combination index number 1
140
159
0
0
0
0
0
00
0
0
0
0
0
1981
488
1280
193
144
1
1462
4838
62
25
2
MNI VAR ENT COR SKE
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Feature combination index number 6
0
0
0
0
488
0
0
0
1
14620
0
1339
206
185
0
0
93
161
1935
2
0
13
4823
67
MNI VAR DIS ENT COR IDM SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
F-re
flect
ion
Oth
er o
bj
Smok
e
Fire
S-re
flect
ion
Predicted class
Figure 9 Continued
Journal of Sensors 11
Feature combination index number 7 Feature combination index number 8
Feature combination index number 9 Feature combination index number 10
89
133
0
0
0
0
0
02
0
0
0
0
0
1888
488
0
0
0
0
488
1341
200
235
1
1459
4857
62
15
5
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1458
0
0
0
1
1462
0
0
0
1
1462
0
0
1324
204
181
0
0
1334
208
181
0
0
1334
205
204
0
0
99
174
1943
0
0
91
150
1921
2
0
18
4810
65
6
0
20
4835
62
0
0
93
172
1941
2
0
22
4812
63
MNI VAR DIS ENT COR IDM SKE KUR STD
MNI VAR DIS ENT IDM SKE KUR STD
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bj
Predicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
MNI VAR DIS ENT IND COR IDM SKE STD
MNI VAR DIS ENT IND COR UNI SKE STD
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
Figure 9 Results of confusion matrix at each feature combination
refined and clearer than the thermal images on the secondrow Smoke fire and their thermal reflections are identifiedand marked in red or orange ellipses
6 Conclusion
The appropriate combination of features was investigated toaccurately classify fire smoke and their thermal reflectionsusing thermal images Gray-scale 14-bit images from a singleinfrared camera were used to extract motion and texture fea-tures by applying a clustering-based autothresholding tech-nique Bayesian classification is performed to probabilisti-cally identify multiple classes during real-time implemen-tation To find the best combination of features a multi-objective genetic algorithm optimization was implementedusing resubstitution and cross-validation errors as objectivefunctions Large-scale fire tests with different fire sources
were conducted to create a range of temperature and smokeconditions to evaluate the feature combinations
Fifteenmotion and texture features were analyzed and theprobability density functions of the features were computedby the maximum likelihood estimation The combinationof multiple features was determined to more accuratelyclassify fire smoke and thermal reflections compared witha single feature In the behavioral solution set where featurecombinations produce less than 7 resubstitution and cross-validation errors COR ENT DIS SKE STD VAR andMNIhad 800 or more occurrence while other features had400 or less occurrence The feature combination of MNIDIS COR SKE and STD produced the highest performancein the classification resulting in 668 and 670 resubstitu-tion and cross-validation errors and 9564 9761 9662and 9345 precision sensitivity F-measure and accuracyrespectively
12 Journal of Sensors
In the near future the classification of fire smoke andtheir thermal reflections will be evaluated on any classifiersand features to increase performanceThe convolution neuralnetwork of deep learning which has recently shown highperformance could be explored as a classifier also model-based image features such as discrete wavelet transform willbe further studied
Appendix
See Figure 9
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this manuscript
Acknowledgments
This work was sponsored by the Office of Naval ResearchGrant no N00014-11-1-0074 scientific office Dr ThomasMcKenna in USA Hwarang-dae Research Institute in Seouland Agency for Defense Development in Daejeon SouthKorea The authors would like to thank Mr Joseph Starr andMr JoshMcNeil for assisting in performing the fire testsTheauthors would also like to thank Rosana K Lee for helpingand supporting this research
References
[1] J-H Kim and B Y Lattimer ldquoReal-time probabilistic classifi-cation of fire and smoke using thermal imagery for intelligentfirefighting robotrdquo Fire Safety Journal vol 72 pp 40ndash49 2015
[2] J W Starr and B Y Lattimer ldquoEvaluation of navigation sensorsin fire smoke environmentsrdquo Fire Technology vol 50 no 6 pp1459ndash1481 2014
[3] J-H Kim B Keller and B Y Lattimer ldquoSensor fusion basedseek-and-find fire algorithm for intelligent firefighting robotrdquoin Proceedings of the IEEEASME International Conference onAdvanced Intelligent Mechatronics (AIM rsquo13) pp 1482ndash1486IEEE Wollongong Australia July 2013
[4] J G McNeil J Starr and B Y Lattimer ldquoAutonomous fire sup-pression using multispectral sensorsrdquo in Proceedings of theIEEEASME International Conference on Advanced IntelligentMechatronics Mechatronics for HumanWellbeing (AIM rsquo13) pp1504ndash1509 Wollongong Austrailia July 2013
[5] J-H Kim J W Starr and B Y Lattimer ldquoFirefighting robotstereo infrared vision and radar sensor fusion for imagingthrough smokerdquo Fire Technology vol 51 no 4 pp 823ndash8452015
[6] RC Luo andK L Su ldquoAutonomous fire-detection systemusingadaptive sensory fusion for intelligent security robotrdquo IEEEASME Transactions on Mechatronics vol 12 no 3 pp 274ndash2812007
[7] M A Jackson and I Robins ldquoGas sensing for fire detectionmeasurements of CO CO
2
H2
O2
and smoke density in Euro-pean standard fire testsrdquo Fire Safety Journal vol 22 no 2 pp181ndash205 1994
[8] B U Toreyin R G Cinbis Y Dedeoglu and A E Cetin ldquoFiredetection in infrared video using wavelet analysisrdquo OpticalEngineering vol 46 no 6 Article ID 067204 2007
[9] B U Toreyin Y Dedeoglu andA E Cetin ldquoWavelet based real-time smoke detection in videordquo in Proceedings of the 13th Euro-pean Signal Processing Conference pp 4ndash8 Antalya TurkeySeptember 2005
[10] T Celik H Demirel H Ozkaramanli and M Uyguroglu ldquoFiredetection using statistical color model in video sequencesrdquoJournal of Visual Communication and Image Representation vol18 no 2 pp 176ndash185 2007
[11] L Merino F Caballero J R Martınez-de-Dios I Maza and AOllero ldquoAn unmanned aircraft system for automatic forest firemonitoring and measurementrdquo Journal of Intelligent amp RoboticSystems vol 65 no 1 pp 533ndash548 2012
[12] Y Wang T W Chua R Chang and N T Pham ldquoReal-timesmoke detection using texture and color featuresrdquo in Proceed-ings of the 21st International Conference on Pattern Recognition(ICPR rsquo12) pp 1727ndash1730 Tsukuba Japan November 2012
[13] G Marbach M Loepfe and T Brupbacher ldquoAn image process-ing technique for fire detection in video imagesrdquo Fire SafetyJournal vol 41 no 4 pp 285ndash289 2006
[14] M I Chacon-Murguia and F J Perez-Vargas ldquoThermal videoanalysis for fire detection using shape regularity and intensitysaturation featuresrdquo in Pattern Recognition J F Martınez-Trinidad J A Carrasco-Ochoa C B-Y Brants and E RHancock Eds vol 6718 of Lecture Notes in Computer Sciencepp 118ndash126 Springer Berlin Germany 2011
[15] W Phillips III M Shah and N D V Lobo ldquoFlame recognitionin videordquo Pattern Recognition Letters vol 23 no 1ndash3 pp 319ndash327 2002
[16] D Han and B Lee ldquoDevelopment of early tunnel fire detectionalgorithm using the image processingrdquo in Advances in VisualComputing pp 39ndash48 Springer Berlin Germany 2006
[17] Y Chunyu F Jun W Jinjun and Z Yongming ldquoVideo firesmoke detection using motion and color featuresrdquo Fire Technol-ogy vol 46 no 3 pp 651ndash663 2010
[18] F Yuan ldquoVideo-based smoke detection with histogram se-quence of LBP and LBPV pyramidsrdquo Fire Safety Journal vol 46no 3 pp 132ndash139 2011
[19] F Lafarge X Descombes and J Zerubia ldquoTextural kernel forSVM classification in remote sensing application to forest firedetection andUrban area extractionrdquo in Proceedings of the IEEEInternational Conference on Image Processing (ICIP rsquo05) pp1096ndash1099 September 2005
[20] F Amon and A Ducharme ldquoImage frequency analysis for test-ing of fire service thermal imaging camerasrdquo Fire Technologyvol 45 no 3 pp 313ndash322 2009
[21] F Amon V Benetis J Kim and A Hamins ldquoDevelopmentof a performance evaluation facility for fire fighting thermalimagersrdquo in Defense and Security pp 244ndash252 2004
[22] F D Maxwell ldquoA portable IR system for observing fire thrusmokerdquo Fire Technology vol 7 no 4 pp 321ndash331 1971
[23] A Barducci D Guzzi P Marcoionni and I Pippi ldquoInfrareddetection of active fires and burnt areas theory and observa-tionsrdquo Infrared Physics amp Technology vol 43 no 3ndash5 pp 119ndash125 2002
[24] C Wang and S Qin ldquoAdaptive detection method of infraredsmall target based on target-background separation via robustprincipal component analysisrdquo Infrared Physics amp Technologyvol 69 pp 123ndash135 2015
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
4 Journal of Sensors
CON = sum119862119894119895(119894 minus 119895)
2
INV = sum119862119894119895
1 +1003816100381610038161003816119894 minus 119895
1003816100381610038161003816
COR = sum(119894119895 + 120583
119894120583119895minus 120583119894minus 120583119895) 119862119894119895
120590119894120590119895
UNI = sum1198622119894119895
IDM = sum
119862119894119895
1 + (119894 minus 119895)2
(3)
4 Object Extraction andBayesian Classification
One of the main characteristics of fire smoke and theirthermal reflections in thermal images is that they are higherin intensity than the background With intensity related totemperature in the thermal image higher temperature objectsappear brighter than the background Hence intensity is aprimary factor for object extraction from the backgroundAssuming that the thermal image histogram has a bimodaldistribution for foreground (ie object) and backgroundthe clustering-based image autothresholding method [34]called Otsumethod can calculate an optimum threshold thatseparates objects and background creating a binary imagewith 0 being the background and 1 being the objects Thebinary images were filtered to remove small regions and holesinside objects through morphological filtering techniquesAfter convoluting the original 14-bit image with the filtered-binary image a final image was obtained that includes theoriginal 14-bit intensities in objects as well as zeroes in thebackground
There are several classification methods commonly usedin supervised machine learning 119896-nearest neighbors (119896NN)decision tree (DT) neural networks (NN) support vectormachine (SVM) and Naıve Bayesian For this study theseclassification methods were analyzed by considering threepoints capability to classify multiple classes such as firesmoke and their thermal reflections less chance of overfit-ting problem because under fire environments there couldbe a number of situations that are not learned or trainedreal-time implementation because firefighting robot needs tomake a decision in real-time otherwise it cannot operate itstask 119896NN is insensitive to outliers but it needs a large amountof memory and expensive computation [35] DT has lowcomputation burden but for the multiclasses classificationit may generate a complicated tree structure and may causeoverfitting problem [35 36] NN shows high performancewhen processing with multidimensions and continuous fea-tures but cannot overcome overfitting problem SVM pro-vides fast computation and the highest accuracy but it cannotbe used for the multilabel classification because it producesbinary results [37] Naıve Bayesian classification is Bayesrsquotheorem-based probabilistic classification and is popular for
pattern recognition applications Although this method haslower accuracy compared with other classifiers and assumesthat each feature is independent it has fast computationrobustness to untrained cases and less chance of overfitting[35] In addition this classification has the capability ofprobabilistic decision making over multiple classes with fastcomputation for real-time implementation In this studyBayesian classification is used for evaluation of each feature
With several given features 1198651 1198652 119865
119902 (motion and
texture features) we can calculate the probability that oneclass 119862
ℎ(fire smoke thermal reflections etc) corresponds
to the candidate 119896 by using a conditional probability 119896119901(119862ℎ|
11986511198652sdot sdot sdot 119865119902) also known as the posterior probability By using
Bayesrsquo theorem it can be written with prior likelihood andevidence as shown in
119896
119901 (119862ℎ| 11986511198652sdot sdot sdot 119865119902)
=
119896
119901 (119862ℎ)119896
119891 (11986511198652sdot sdot sdot 119865119902| 119862ℎ)
sum119862ℎ
119896119901 (119862ℎ)119896
119891 (11986511198652sdot sdot sdot 119865119902| 119862ℎ)
(4)
where 119896119901(119862ℎ) is the prior probability meaning it represents
candidate 119896 probability to be 119862ℎand can be calculated by
number of samples of class 119862ℎdivided by the total number
of samples 119896119891(11986511198652sdot sdot sdot 119865119902| 119862ℎ) is the likelihood function
and the denominator of (4) is the evidence that plays asa normalizing constant by the summation of productionbetween the prior and likelihood at each class By applying theconditional independence assumption the likelihood func-tion can be rewritten by
119891 (11986511198652sdot sdot sdot 119865119902| 119862ℎ) =
119902
prod
119894=1
119891 (119865119894| 119862ℎ) (5)
The conditional probability density function 119891(119865119894| 119862ℎ) sim
119873(120583119865119894|119862ℎ
Σ119865119894|119862ℎ
) can be described as
119891 (119865119894| 119862ℎ) =
1
radic2120587Σ119865119894|119862ℎ
119890minus(12)(119865119894minus120583119865119894|119862ℎ
)
119879Σ
minus1
119865119894|119862ℎ(119865119894minus120583119865119894|119862ℎ
)
(6)
where
120583119865119894|119862ℎ
= 120583119865119894+ Σ119879
119862ℎ119865119894
Σminus1
119862ℎ119862ℎ
(119862ℎminus 120583119862ℎ)
Σ119865119894|119862ℎ
= Σ119865119894119865119894minus Σ119879
119862ℎ119865119894
Σminus1
119862ℎ119862ℎ
Σ119862ℎ119865119894
(7)
As shown in Table 2 Gaussian parameters for fifteen featureswith respect to smoke smoke thermal reflection fire andfire thermal reflection were estimated by using the maximumlikelihood estimation [38] Probability density distributionsfor the entire features are illustrated in Figure 3 With (5) theevidence and then the posterior probability of each class werecalculated By applying the maximum priority decision rulein (8) the Bayesian classification was used to predict the classand probability of each candidate in the scene
class = argmax119862ℎ
(119901 (119862ℎ| 11986511198652sdot sdot sdot 119865119902))
prob = max (119901 (119862ℎ| 11986511198652sdot sdot sdot 119865119902))
(8)
Journal of Sensors 5
Table 2 Gaussian parameters
Smoke Smoke-reflection Fire Fire-reflection120583 Σ 120583 Σ 120583 Σ 120583 Σ
MNI minus12665119864 + 04 69008119864 + 02 minus13383119864 + 04 35543119864 + 02 minus59714119864 + 03 16050119864 + 03 minus71399119864 + 03 46070119864 + 02
VAR 47578119864 + 05 44154119864 + 05 44012119864 + 04 41227119864 + 04 10501119864 + 07 22343119864 + 06 13300119864 + 06 88178119864 + 05
NOP 10733119864 + 03 42287119864 + 03 32220119864 + 01 13676119864 + 02 22174119864 + 02 17105119864 + 03 59575119864 + 01 23911119864 + 02
STD 62146119864 + 02 29931119864 + 02 18314119864 + 02 10236119864 + 02 32170119864 + 03 38930119864 + 02 10534119864 + 03 46992119864 + 02
SKE 11672119864 minus 01 58208119864 minus 01 minus94009119864 minus 02 62857119864 minus 01 22230119864 minus 02 52287119864 minus 01 22385119864 minus 01 71380119864 minus 01
KUR 30045119864 + 00 20213119864 + 00 36553119864 + 00 16131119864 + 00 22283119864 + 00 11517119864 + 00 34848119864 + 00 11912119864 + 00
OFVN 28832119864 + 04 14156119864 + 04 11848119864 + 03 87345119864 + 02 11257119864 + 04 13722119864 + 04 29978119864 + 03 22034119864 + 03
OFVMM 86830119864 + 01 56259119864 + 01 11687119864 + 02 79444119864 + 01 17598119864 + 02 29308119864 + 02 12641119864 + 02 67234119864 + 01
DIS 71791119864 minus 02 17966119864 minus 02 28045119864 minus 02 12245119864 minus 02 10937119864 minus 01 86515119864 minus 02 47308119864 minus 02 28550119864 minus 02
ENT 41949119864 minus 01 93683119864 minus 02 19576119864 minus 01 86106119864 minus 02 33681119864 minus 01 19107119864 minus 01 17115119864 minus 01 99702119864 minus 02
CON 89384119864 minus 01 36871119864 minus 01 12402119864 minus 01 65921119864 minus 02 92956119864 minus 01 69179119864 minus 01 39683119864 minus 01 26996119864 minus 01
IND 98588119864 minus 01 63845119864 minus 03 99646119864 minus 01 15141119864 minus 03 96334119864 minus 01 37231119864 minus 02 98771119864 minus 01 79275119864 minus 03
COR 96636119864 minus 01 31625119864 minus 02 88533119864 minus 01 43467119864 minus 02 89406119864 minus 01 30935119864 minus 02 88417119864 minus 01 57861119864 minus 02
UNI 51044119864 minus 01 20109119864 minus 01 95222119864 minus 01 31669119864 minus 02 67906119864 minus 01 26718119864 minus 01 85305119864 minus 01 10599119864 minus 01
IDM 60061119864 minus 01 19160119864 minus 01 97533119864 minus 01 16844119864 minus 02 77850119864 minus 01 24321119864 minus 01 92043119864 minus 01 58631119864 minus 02
Table 3 The object numbers of smoke smoke-reflection fire fire-reflection and other hot objects classes
Type Total Smoke Smoke-reflection Fire Fire-reflection Other hot objectsNumber of objects 10775 5190 1445 1464 489 2187
Figure 3 shows probability density distribution of eachclass using the Gaussian parameters of Table 2 Gaussiandistribution for classes in Figure 3 shows how fire fire-reflection smoke and smoke-reflection are distributed by thefifteen features Some features split out the distribution of thefour classes while others cause overlap For example MNIbest describes a well split out case of the classes althoughsmoke and its reflection and fire and its reflection do overlapSKE shows the worst case in which all classes overlap makingit impossible to distinguish any of the four classes
5 Result and Discussion
The accuracy in classifying fire objects was analyzed usingdata from a series of large-scale tests in the facility [1]using actual fires up to 75 kW Fires included latex foamwood cribs and propane gas fires from a sand burner Thesedifferent types of fires produced a range of temperature andsmoke conditions Latex foam fires produced lower tempera-ture conditions but dense low visibility smoke Converselypropane gas fires produced higher gas temperatures andlight smoke Wood crib fires resulted in smoke and gastemperatures between those of latex foam and propane gasfires however these fires resulted in sparks created from theburning wood Thermal images were collected by driving awheeled mobile robot through the setup during a fire test Atotal of 10775 objects were collected from the experimentsand categorized as either smoke smoke-reflection fire fire-reflection or other hot objects in order to be served as cluesto lead the firefighting robot to navigate toward the fire source
outside the FOV In addition as each object has sixteen cor-responding data points (fifteen features and a class) the totalnumber of data points used in this paper is 172400The num-bers of each object in this experiment are shown in Table 3
Two types of error criterions (resubstitution and 119896-foldcross-validation errors [39]) were used to measure howeach feature accurately performs in the classification Resub-stitution error takes the entire dataset to compare the actualclasses with the predicted classes by the Bayesian classifica-tion in order to examine how well the actual and predictedclasses match each other When this criterion is used aloneto enhance accuracy the classification can be overfitted tothe training dataset Cross-validation error is advantageousto detect and prevent from overfitting Instead of using theentire dataset cross-validation randomly selects and splitsthe dataset into 119896 partitions of approximately equal size(119896 = 10) to estimate a mean error by comparing betweenthe randomly selected partition and trained results of theremaining partitions
51 Single Feature Performance The performance results ofeach feature are shown in Table 4 The first-order statisticaltexture features MNI VAR and STD produced the lowesterrors while NOP SKE and KUR show the highest Theseresults show that MNI and VAR are beneficial to distinguishfire smoke and thermal reflections while motion features arenot As NOP shows the highest error OFVMM one of themotion features shows the second highest errors comparedwith the other features This is in part attributed to thedynamic motion of the robot ENT and COR second-order
6 Journal of Sensors
0 5 10VAR
Prob
abili
ty d
ensit
y
SmokeS-reflection
FireF-reflection
times10minus4
times10minus4
times104
times10minus7
times106
times10minus3
2468
1012
Prob
abili
ty d
ensit
y
minus10000 minus8000 minus6000 minus4000minus12000MNI
0
5
10
Prob
abili
ty d
ensit
y
0
05
1
Prob
abili
ty d
ensit
y
5000 100000OFVN
0
20
40
Prob
abili
ty d
ensit
y
01 02 03 040DIS
500 1000 1500 20000OFVMM
0
0005
001
Prob
abili
ty d
ensit
y0
5
Prob
abili
ty d
ensit
y
2 4 60NOP
50100150200250
Prob
abili
ty d
ensit
y
09 095 1085IND
1 15 2 2505CON
2
4
6
Prob
abili
ty d
ensit
y
02 04 06 08 10ENT
0
5
Prob
abili
ty d
ensit
y
2468
1012
Prob
abili
ty d
ensit
y
06 08 104COR
minus2 0 2 4 6minus4SKE
0
05
1
Prob
abili
ty d
ensit
y
01
02
03
Prob
abili
ty d
ensit
y
5 10 15 200KUT
1000 2000 3000 40000STD
0
2
4
08 10604IDM
510152025
Prob
abili
ty d
ensit
y
02 04 06 08 10UNI
0
10
20
Prob
abili
ty d
ensit
y
SmokeS-reflection
FireF-reflection
SmokeS-reflection
FireF-reflection
times10minus3
Figure 3 Probability density distributions of each feature
statistical texture features show42sim45error which is higherthan the other second-order features
52 Multiple Feature Combination Performance The errorresults in Table 4 demonstrate that a single feature cannotaccurately classify fire smoke and thermal reflections Thuspossible combination ofmultiple features was considered andanalyzed to find the best combination of the featuresThe totalnumber of all possible combinations that have two or morefeatures is
119873total =119898
sum
119911=2
(
119898
119911
) (9)
where 119898 refers to the total number of features (ie 119898 = 15)and 119911 is the number of features in the combination Basedon all possible combination the multiobjective genetic algo-rithm optimization [40] in the global optimization toolbox ofMATLAB was used to find the best combination of featuresthat has the highest performance in the classification Theobjective functions in the optimization resubstitution and119896-fold cross-validation errors [39] were used to measurehow accurately different feature combinations perform in theclassification
Figure 4 contains a plot of the error associated with themost promising feature combinations The behavioral solu-tion set is defined as feature combinations with less than 7
Journal of Sensors 7
Table 4 Performance of each feature
Resubstitution error () Cross-validation error ()MNI 237 238VAR 243 243NOP 721 720STD 232 232SKE 527 528KUR 506 506OFVN 395 400OFVMM 586 586DIS 290 290ENT 446 446CON 285 285IND 301 301COR 412 411UNI 345 345IDM 377 377
Behavioral region
Behavioral solution setGeneral set
66 67 68 69 7 71 72 7365Resubstitute error ()
65
66
67
68
69
7
71
72
73
Cros
s-va
lidat
ion
erro
r (
)
Figure 4Multiobjective optimization result showing the general setand behavioral solution set (colored region)
error for both objective functions while the general set refersto all other possible feature combinations The behavioralsolution set contains 00061 of all possible feature combi-nations
The occurrence probability of features in the behavioralsolution set is illustrated in Figure 5 In the behavioralsolution set the first-order statistic texture features MNI andSKE always exist while OFVN NOP and OFVMM featuresdo not Both the first-order statistical texture features STDand VAR and the second-order statistical texture featuresCOR ENT andDIS showahigher occurrence comparedwiththe other first- and second-order texture features while KURIDMUNI IND andCON show lower occurrence Note thatdue to the robotrsquos dynamical motion motion features werenot successful and even not included in the top 10 featurecombinations of the behavioral set
1000800
00900
1000200
0000
800900
100200
900200
400
MNIVARNOPSTDSKE
KUROFVN
OFMMDIS
ENTCONINDCORUNIIDM
200 400 600 800 1000 120000Occurrence probability ()
Figure 5 Occurrence analysis of the features in the behavioralregion
The top features based on the probability occurrence inFigure 5 are COR ENT DIS SKE STD VAR and MNIHowever the combination of these seven features does notresult in the best solution for classification Table 5 containsthe classification performance of the combination of featuresin the behavioral solution set In order to evaluate the per-formance of each feature combination various performancemeasures have been used such as precision sensitivity F-measure and accuracy Precision measures the fraction ofpositive instances from the group that the classifier predictedto be positive and recall measures the fraction of positiveexamples from the positive group of the actual class and [35]F-measure is the harmonic mean and accuracy is the pro-portion of true resultsThesemeasures can bemathematicallydefined as
Precision = TPTP + FP
Sensitivity = TPTP + FN
119865-Measure =2 (Sensitivity sdot Precision)Sensitivity + Precision
Accuracy = TPTP + FP + FN
(10)
where TP is correctly classified positive cases FP is incor-rectly classified negative cases and FN is incorrectly classifiedpositive cases For the performance measurement confusionmatrixes were created as described in Appendix and appliedinto (10) In the precision index number 1 combinationshows the highest performance in the behavioral solutionset while index number 7 combination shows the lowestIn the sensitivity index number 7 combination records thehighest results while index number 4 does the lowest In theF-measure and accuracy index number 2 combination showsthe highest record while index number 4 does the lowestBased on the confusion matrixes most of misclassificationoccurs in the classification of smoke smoke-reflection andother hot objects because during small fire texture patternsof these classes were diminished and the intensity was too low
8 Journal of Sensors
Table 5 Results of error case and performance at each feature combination in the behavioral solution set (Resu means resubstitution andCros refers to cross-validation)
Index Combination of features Error () Case Performance ()Resu Cros TP FP FN Precision Sensitivity 119865-measure Accuracy
1 MNI VAR ENT COR SKE 690 689 10049 402 324 9615 9688 9651 93262 MNI DIS COR SKE STD 668 670 10069 459 247 9564 9761 9662 93453 MNI ENT COR SKE STD 676 675 10061 447 267 9575 9741 9657 9337
4 MNI VAR DIS ENT CON CORSKE STD 698 690 9980 454 341 9565 9670 9617 9262
5 MNI VAR DIS ENT COR UNISKE STD 689 690 10037 453 285 9568 9724 9645 9315
6 MNI VAR DIS ENT COR IDMSKE STD 677 679 10047 461 267 9561 9741 9650 9324
7 MNI VAR DIS ENT IDM SKEKUR STD 696 694 10033 505 237 9521 9769 9643 9311
8 MNI VAR DIS ENT IND CORUNI SKE STD 695 699 10029 451 295 9570 9714 9641 9308
9 MNI VAR DIS ENT IND CORIDM SKE STD 688 690 10035 457 283 9564 9726 9644 9313
10 MNI VAR DIS ENT COR IDMSKE KUR STD 689 691 10036 478 261 9545 9747 9645 9314
PrecisionSensitivity
F-measureAccuracy
90919293949596979899
Perfo
rman
ce (
)
2 3 4 5 6 7 8 9 101Feature combination index
Figure 6 Results of performance at each feature combination in thebehavioral solution set
to distinguish The best solution was determined to be indexnumber 2 combination of MNI DIS COR SKE and STDwhich has the lowest of resubstitution and cross-validationerrors 668 and 670 respectively This combinationincludes all of the top features based on the probability occur-rence except ENT and VAR The four performance results ateach feature combination in the behavioral solution set areshown in Figure 6 where the highest results aremarked in redcircles and the lowest in green-dot circles Sensitivity appearshigher than precision at each feature combination becauseFPs are larger than FNs in the confusion matrix Particularlyindex number 7 has the biggest difference between FP and FNresulting in the highest sensitivity and lowest precision Thesummation of FP and FN in index number 4 is the highestin the behavioral solution set resulting in the lowest accuracywhile index number 2 has the lowest summation of FP andFN providing the highest accuracy
This study investigated a wide range of features fromlong-wavelength infrared camera images analyzed normaldistributions of fifteen features with respect to the classes ofsmoke fire and their thermal reflections and discovered thehighest performing feature combination by examining singlefeatures and multiple feature combinations As a result theproposed feature combination of MNI DIS COR SKE andSTD increases the performance compared with the previousstudy [1] which used MNI VAR ENT and IDM As shownin Figure 7 the errors are reduced by 286 and 268 resub-stitution and cross-validation errors and performances areincreased by 290 158 020 and 285 accuracy F-measure sensitivity and precision respectively
Figure 8 shows original visual and thermal images withthe robot at three different locations start point hallwayentrance and room entrance described in the experimentalfacility Each row relates to a series of images from the robotat three locations The first row contains visible images ofthe robot view As seen in the visible image at start pointfurther information regarding the hallway is limited due toshadowing of the light The image at hallway entrance showsa smoke layer in the upper portion of the hallway due to afire inside the roomThe image at the room entrance displaysa wood crib fire with sparks Because of soot and relativedifference in brightness the background is shown darker andthus limiting information on the background around the fire
Thermal infrared images are displayed in the second rowto show information that RGB camera cannot provide infire environments Unlike visual image at start point thatis obscured due to shadowing the presence of smoke andits thermal reflections on the ventilation hood can be obvi-ously perceived The red boxes on thermal images indicateobjects extracted through the adaptive object extraction withoptical flows and identification numbers In spite of dense
Journal of Sensors 9
9279
9741
9504
9055
954
938
9564
9761
9662
9345
668
670
Precision
Sensitivity
F-measure
Accuracy
Resubstitution error
Cross-validation error
The proposedThe previous
2000 4000 6000 8000 10000000()
Figure 7 Result comparison between the previous and the proposed studies
① Start point ② Hallway entrance ③ Room entrance
IR im
age
class
ifica
tion
resu
lts
Visib
le im
age
IR im
age
imag
e ana
lysis
Figure 8 Original visible and IR images with image analysis and classification results at different locations in the test setup of actual fires
smoke-filled and low visibility environments thermal imagescan generate the images of smoke and fire as well as back-ground information that is otherwise not visible throughvisual imaging
On the third row class labels and posterior probabilitiesof each candidate are displayed at the center of candidateROI as a result of Bayesian classification Using enhancedimage processing techniques the thermal images can bemore
10 Journal of Sensors
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
2
0
19
4820
64
Feature combination index number 2
Feature combination index number 3 Feature combination index number 4
Feature combination index number 5
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1462
0
0
0
1
1462
0
0
0
1
1457
0
0
0
1
1457
0
0
1365
193
197
0
0
1343
181
196
0
0
1325
205
181
0
0
1270
207
181
0
0
53
167
1929
0
0
145
166
1943
0
0
101
165
1942
7
0
49
4844
62
2
0
30
4817
63
0
0
53
165
1929
7
0
27
4830
61
MNI DIS COR SKE STD
MNI ENT COR SKE STD MNI VAR DIS ENT CON COR SKE STD
MNI VAR DIS ENT COR UNI SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bjPredicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Feature combination index number 1
140
159
0
0
0
0
0
00
0
0
0
0
0
1981
488
1280
193
144
1
1462
4838
62
25
2
MNI VAR ENT COR SKE
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Feature combination index number 6
0
0
0
0
488
0
0
0
1
14620
0
1339
206
185
0
0
93
161
1935
2
0
13
4823
67
MNI VAR DIS ENT COR IDM SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
F-re
flect
ion
Oth
er o
bj
Smok
e
Fire
S-re
flect
ion
Predicted class
Figure 9 Continued
Journal of Sensors 11
Feature combination index number 7 Feature combination index number 8
Feature combination index number 9 Feature combination index number 10
89
133
0
0
0
0
0
02
0
0
0
0
0
1888
488
0
0
0
0
488
1341
200
235
1
1459
4857
62
15
5
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1458
0
0
0
1
1462
0
0
0
1
1462
0
0
1324
204
181
0
0
1334
208
181
0
0
1334
205
204
0
0
99
174
1943
0
0
91
150
1921
2
0
18
4810
65
6
0
20
4835
62
0
0
93
172
1941
2
0
22
4812
63
MNI VAR DIS ENT COR IDM SKE KUR STD
MNI VAR DIS ENT IDM SKE KUR STD
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bj
Predicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
MNI VAR DIS ENT IND COR IDM SKE STD
MNI VAR DIS ENT IND COR UNI SKE STD
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
Figure 9 Results of confusion matrix at each feature combination
refined and clearer than the thermal images on the secondrow Smoke fire and their thermal reflections are identifiedand marked in red or orange ellipses
6 Conclusion
The appropriate combination of features was investigated toaccurately classify fire smoke and their thermal reflectionsusing thermal images Gray-scale 14-bit images from a singleinfrared camera were used to extract motion and texture fea-tures by applying a clustering-based autothresholding tech-nique Bayesian classification is performed to probabilisti-cally identify multiple classes during real-time implemen-tation To find the best combination of features a multi-objective genetic algorithm optimization was implementedusing resubstitution and cross-validation errors as objectivefunctions Large-scale fire tests with different fire sources
were conducted to create a range of temperature and smokeconditions to evaluate the feature combinations
Fifteenmotion and texture features were analyzed and theprobability density functions of the features were computedby the maximum likelihood estimation The combinationof multiple features was determined to more accuratelyclassify fire smoke and thermal reflections compared witha single feature In the behavioral solution set where featurecombinations produce less than 7 resubstitution and cross-validation errors COR ENT DIS SKE STD VAR andMNIhad 800 or more occurrence while other features had400 or less occurrence The feature combination of MNIDIS COR SKE and STD produced the highest performancein the classification resulting in 668 and 670 resubstitu-tion and cross-validation errors and 9564 9761 9662and 9345 precision sensitivity F-measure and accuracyrespectively
12 Journal of Sensors
In the near future the classification of fire smoke andtheir thermal reflections will be evaluated on any classifiersand features to increase performanceThe convolution neuralnetwork of deep learning which has recently shown highperformance could be explored as a classifier also model-based image features such as discrete wavelet transform willbe further studied
Appendix
See Figure 9
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this manuscript
Acknowledgments
This work was sponsored by the Office of Naval ResearchGrant no N00014-11-1-0074 scientific office Dr ThomasMcKenna in USA Hwarang-dae Research Institute in Seouland Agency for Defense Development in Daejeon SouthKorea The authors would like to thank Mr Joseph Starr andMr JoshMcNeil for assisting in performing the fire testsTheauthors would also like to thank Rosana K Lee for helpingand supporting this research
References
[1] J-H Kim and B Y Lattimer ldquoReal-time probabilistic classifi-cation of fire and smoke using thermal imagery for intelligentfirefighting robotrdquo Fire Safety Journal vol 72 pp 40ndash49 2015
[2] J W Starr and B Y Lattimer ldquoEvaluation of navigation sensorsin fire smoke environmentsrdquo Fire Technology vol 50 no 6 pp1459ndash1481 2014
[3] J-H Kim B Keller and B Y Lattimer ldquoSensor fusion basedseek-and-find fire algorithm for intelligent firefighting robotrdquoin Proceedings of the IEEEASME International Conference onAdvanced Intelligent Mechatronics (AIM rsquo13) pp 1482ndash1486IEEE Wollongong Australia July 2013
[4] J G McNeil J Starr and B Y Lattimer ldquoAutonomous fire sup-pression using multispectral sensorsrdquo in Proceedings of theIEEEASME International Conference on Advanced IntelligentMechatronics Mechatronics for HumanWellbeing (AIM rsquo13) pp1504ndash1509 Wollongong Austrailia July 2013
[5] J-H Kim J W Starr and B Y Lattimer ldquoFirefighting robotstereo infrared vision and radar sensor fusion for imagingthrough smokerdquo Fire Technology vol 51 no 4 pp 823ndash8452015
[6] RC Luo andK L Su ldquoAutonomous fire-detection systemusingadaptive sensory fusion for intelligent security robotrdquo IEEEASME Transactions on Mechatronics vol 12 no 3 pp 274ndash2812007
[7] M A Jackson and I Robins ldquoGas sensing for fire detectionmeasurements of CO CO
2
H2
O2
and smoke density in Euro-pean standard fire testsrdquo Fire Safety Journal vol 22 no 2 pp181ndash205 1994
[8] B U Toreyin R G Cinbis Y Dedeoglu and A E Cetin ldquoFiredetection in infrared video using wavelet analysisrdquo OpticalEngineering vol 46 no 6 Article ID 067204 2007
[9] B U Toreyin Y Dedeoglu andA E Cetin ldquoWavelet based real-time smoke detection in videordquo in Proceedings of the 13th Euro-pean Signal Processing Conference pp 4ndash8 Antalya TurkeySeptember 2005
[10] T Celik H Demirel H Ozkaramanli and M Uyguroglu ldquoFiredetection using statistical color model in video sequencesrdquoJournal of Visual Communication and Image Representation vol18 no 2 pp 176ndash185 2007
[11] L Merino F Caballero J R Martınez-de-Dios I Maza and AOllero ldquoAn unmanned aircraft system for automatic forest firemonitoring and measurementrdquo Journal of Intelligent amp RoboticSystems vol 65 no 1 pp 533ndash548 2012
[12] Y Wang T W Chua R Chang and N T Pham ldquoReal-timesmoke detection using texture and color featuresrdquo in Proceed-ings of the 21st International Conference on Pattern Recognition(ICPR rsquo12) pp 1727ndash1730 Tsukuba Japan November 2012
[13] G Marbach M Loepfe and T Brupbacher ldquoAn image process-ing technique for fire detection in video imagesrdquo Fire SafetyJournal vol 41 no 4 pp 285ndash289 2006
[14] M I Chacon-Murguia and F J Perez-Vargas ldquoThermal videoanalysis for fire detection using shape regularity and intensitysaturation featuresrdquo in Pattern Recognition J F Martınez-Trinidad J A Carrasco-Ochoa C B-Y Brants and E RHancock Eds vol 6718 of Lecture Notes in Computer Sciencepp 118ndash126 Springer Berlin Germany 2011
[15] W Phillips III M Shah and N D V Lobo ldquoFlame recognitionin videordquo Pattern Recognition Letters vol 23 no 1ndash3 pp 319ndash327 2002
[16] D Han and B Lee ldquoDevelopment of early tunnel fire detectionalgorithm using the image processingrdquo in Advances in VisualComputing pp 39ndash48 Springer Berlin Germany 2006
[17] Y Chunyu F Jun W Jinjun and Z Yongming ldquoVideo firesmoke detection using motion and color featuresrdquo Fire Technol-ogy vol 46 no 3 pp 651ndash663 2010
[18] F Yuan ldquoVideo-based smoke detection with histogram se-quence of LBP and LBPV pyramidsrdquo Fire Safety Journal vol 46no 3 pp 132ndash139 2011
[19] F Lafarge X Descombes and J Zerubia ldquoTextural kernel forSVM classification in remote sensing application to forest firedetection andUrban area extractionrdquo in Proceedings of the IEEEInternational Conference on Image Processing (ICIP rsquo05) pp1096ndash1099 September 2005
[20] F Amon and A Ducharme ldquoImage frequency analysis for test-ing of fire service thermal imaging camerasrdquo Fire Technologyvol 45 no 3 pp 313ndash322 2009
[21] F Amon V Benetis J Kim and A Hamins ldquoDevelopmentof a performance evaluation facility for fire fighting thermalimagersrdquo in Defense and Security pp 244ndash252 2004
[22] F D Maxwell ldquoA portable IR system for observing fire thrusmokerdquo Fire Technology vol 7 no 4 pp 321ndash331 1971
[23] A Barducci D Guzzi P Marcoionni and I Pippi ldquoInfrareddetection of active fires and burnt areas theory and observa-tionsrdquo Infrared Physics amp Technology vol 43 no 3ndash5 pp 119ndash125 2002
[24] C Wang and S Qin ldquoAdaptive detection method of infraredsmall target based on target-background separation via robustprincipal component analysisrdquo Infrared Physics amp Technologyvol 69 pp 123ndash135 2015
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
Journal of Sensors 5
Table 2 Gaussian parameters
Smoke Smoke-reflection Fire Fire-reflection120583 Σ 120583 Σ 120583 Σ 120583 Σ
MNI minus12665119864 + 04 69008119864 + 02 minus13383119864 + 04 35543119864 + 02 minus59714119864 + 03 16050119864 + 03 minus71399119864 + 03 46070119864 + 02
VAR 47578119864 + 05 44154119864 + 05 44012119864 + 04 41227119864 + 04 10501119864 + 07 22343119864 + 06 13300119864 + 06 88178119864 + 05
NOP 10733119864 + 03 42287119864 + 03 32220119864 + 01 13676119864 + 02 22174119864 + 02 17105119864 + 03 59575119864 + 01 23911119864 + 02
STD 62146119864 + 02 29931119864 + 02 18314119864 + 02 10236119864 + 02 32170119864 + 03 38930119864 + 02 10534119864 + 03 46992119864 + 02
SKE 11672119864 minus 01 58208119864 minus 01 minus94009119864 minus 02 62857119864 minus 01 22230119864 minus 02 52287119864 minus 01 22385119864 minus 01 71380119864 minus 01
KUR 30045119864 + 00 20213119864 + 00 36553119864 + 00 16131119864 + 00 22283119864 + 00 11517119864 + 00 34848119864 + 00 11912119864 + 00
OFVN 28832119864 + 04 14156119864 + 04 11848119864 + 03 87345119864 + 02 11257119864 + 04 13722119864 + 04 29978119864 + 03 22034119864 + 03
OFVMM 86830119864 + 01 56259119864 + 01 11687119864 + 02 79444119864 + 01 17598119864 + 02 29308119864 + 02 12641119864 + 02 67234119864 + 01
DIS 71791119864 minus 02 17966119864 minus 02 28045119864 minus 02 12245119864 minus 02 10937119864 minus 01 86515119864 minus 02 47308119864 minus 02 28550119864 minus 02
ENT 41949119864 minus 01 93683119864 minus 02 19576119864 minus 01 86106119864 minus 02 33681119864 minus 01 19107119864 minus 01 17115119864 minus 01 99702119864 minus 02
CON 89384119864 minus 01 36871119864 minus 01 12402119864 minus 01 65921119864 minus 02 92956119864 minus 01 69179119864 minus 01 39683119864 minus 01 26996119864 minus 01
IND 98588119864 minus 01 63845119864 minus 03 99646119864 minus 01 15141119864 minus 03 96334119864 minus 01 37231119864 minus 02 98771119864 minus 01 79275119864 minus 03
COR 96636119864 minus 01 31625119864 minus 02 88533119864 minus 01 43467119864 minus 02 89406119864 minus 01 30935119864 minus 02 88417119864 minus 01 57861119864 minus 02
UNI 51044119864 minus 01 20109119864 minus 01 95222119864 minus 01 31669119864 minus 02 67906119864 minus 01 26718119864 minus 01 85305119864 minus 01 10599119864 minus 01
IDM 60061119864 minus 01 19160119864 minus 01 97533119864 minus 01 16844119864 minus 02 77850119864 minus 01 24321119864 minus 01 92043119864 minus 01 58631119864 minus 02
Table 3 The object numbers of smoke smoke-reflection fire fire-reflection and other hot objects classes
Type Total Smoke Smoke-reflection Fire Fire-reflection Other hot objectsNumber of objects 10775 5190 1445 1464 489 2187
Figure 3 shows probability density distribution of eachclass using the Gaussian parameters of Table 2 Gaussiandistribution for classes in Figure 3 shows how fire fire-reflection smoke and smoke-reflection are distributed by thefifteen features Some features split out the distribution of thefour classes while others cause overlap For example MNIbest describes a well split out case of the classes althoughsmoke and its reflection and fire and its reflection do overlapSKE shows the worst case in which all classes overlap makingit impossible to distinguish any of the four classes
5 Result and Discussion
The accuracy in classifying fire objects was analyzed usingdata from a series of large-scale tests in the facility [1]using actual fires up to 75 kW Fires included latex foamwood cribs and propane gas fires from a sand burner Thesedifferent types of fires produced a range of temperature andsmoke conditions Latex foam fires produced lower tempera-ture conditions but dense low visibility smoke Converselypropane gas fires produced higher gas temperatures andlight smoke Wood crib fires resulted in smoke and gastemperatures between those of latex foam and propane gasfires however these fires resulted in sparks created from theburning wood Thermal images were collected by driving awheeled mobile robot through the setup during a fire test Atotal of 10775 objects were collected from the experimentsand categorized as either smoke smoke-reflection fire fire-reflection or other hot objects in order to be served as cluesto lead the firefighting robot to navigate toward the fire source
outside the FOV In addition as each object has sixteen cor-responding data points (fifteen features and a class) the totalnumber of data points used in this paper is 172400The num-bers of each object in this experiment are shown in Table 3
Two types of error criterions (resubstitution and 119896-foldcross-validation errors [39]) were used to measure howeach feature accurately performs in the classification Resub-stitution error takes the entire dataset to compare the actualclasses with the predicted classes by the Bayesian classifica-tion in order to examine how well the actual and predictedclasses match each other When this criterion is used aloneto enhance accuracy the classification can be overfitted tothe training dataset Cross-validation error is advantageousto detect and prevent from overfitting Instead of using theentire dataset cross-validation randomly selects and splitsthe dataset into 119896 partitions of approximately equal size(119896 = 10) to estimate a mean error by comparing betweenthe randomly selected partition and trained results of theremaining partitions
51 Single Feature Performance The performance results ofeach feature are shown in Table 4 The first-order statisticaltexture features MNI VAR and STD produced the lowesterrors while NOP SKE and KUR show the highest Theseresults show that MNI and VAR are beneficial to distinguishfire smoke and thermal reflections while motion features arenot As NOP shows the highest error OFVMM one of themotion features shows the second highest errors comparedwith the other features This is in part attributed to thedynamic motion of the robot ENT and COR second-order
6 Journal of Sensors
0 5 10VAR
Prob
abili
ty d
ensit
y
SmokeS-reflection
FireF-reflection
times10minus4
times10minus4
times104
times10minus7
times106
times10minus3
2468
1012
Prob
abili
ty d
ensit
y
minus10000 minus8000 minus6000 minus4000minus12000MNI
0
5
10
Prob
abili
ty d
ensit
y
0
05
1
Prob
abili
ty d
ensit
y
5000 100000OFVN
0
20
40
Prob
abili
ty d
ensit
y
01 02 03 040DIS
500 1000 1500 20000OFVMM
0
0005
001
Prob
abili
ty d
ensit
y0
5
Prob
abili
ty d
ensit
y
2 4 60NOP
50100150200250
Prob
abili
ty d
ensit
y
09 095 1085IND
1 15 2 2505CON
2
4
6
Prob
abili
ty d
ensit
y
02 04 06 08 10ENT
0
5
Prob
abili
ty d
ensit
y
2468
1012
Prob
abili
ty d
ensit
y
06 08 104COR
minus2 0 2 4 6minus4SKE
0
05
1
Prob
abili
ty d
ensit
y
01
02
03
Prob
abili
ty d
ensit
y
5 10 15 200KUT
1000 2000 3000 40000STD
0
2
4
08 10604IDM
510152025
Prob
abili
ty d
ensit
y
02 04 06 08 10UNI
0
10
20
Prob
abili
ty d
ensit
y
SmokeS-reflection
FireF-reflection
SmokeS-reflection
FireF-reflection
times10minus3
Figure 3 Probability density distributions of each feature
statistical texture features show42sim45error which is higherthan the other second-order features
52 Multiple Feature Combination Performance The errorresults in Table 4 demonstrate that a single feature cannotaccurately classify fire smoke and thermal reflections Thuspossible combination ofmultiple features was considered andanalyzed to find the best combination of the featuresThe totalnumber of all possible combinations that have two or morefeatures is
119873total =119898
sum
119911=2
(
119898
119911
) (9)
where 119898 refers to the total number of features (ie 119898 = 15)and 119911 is the number of features in the combination Basedon all possible combination the multiobjective genetic algo-rithm optimization [40] in the global optimization toolbox ofMATLAB was used to find the best combination of featuresthat has the highest performance in the classification Theobjective functions in the optimization resubstitution and119896-fold cross-validation errors [39] were used to measurehow accurately different feature combinations perform in theclassification
Figure 4 contains a plot of the error associated with themost promising feature combinations The behavioral solu-tion set is defined as feature combinations with less than 7
Journal of Sensors 7
Table 4 Performance of each feature
Resubstitution error () Cross-validation error ()MNI 237 238VAR 243 243NOP 721 720STD 232 232SKE 527 528KUR 506 506OFVN 395 400OFVMM 586 586DIS 290 290ENT 446 446CON 285 285IND 301 301COR 412 411UNI 345 345IDM 377 377
Behavioral region
Behavioral solution setGeneral set
66 67 68 69 7 71 72 7365Resubstitute error ()
65
66
67
68
69
7
71
72
73
Cros
s-va
lidat
ion
erro
r (
)
Figure 4Multiobjective optimization result showing the general setand behavioral solution set (colored region)
error for both objective functions while the general set refersto all other possible feature combinations The behavioralsolution set contains 00061 of all possible feature combi-nations
The occurrence probability of features in the behavioralsolution set is illustrated in Figure 5 In the behavioralsolution set the first-order statistic texture features MNI andSKE always exist while OFVN NOP and OFVMM featuresdo not Both the first-order statistical texture features STDand VAR and the second-order statistical texture featuresCOR ENT andDIS showahigher occurrence comparedwiththe other first- and second-order texture features while KURIDMUNI IND andCON show lower occurrence Note thatdue to the robotrsquos dynamical motion motion features werenot successful and even not included in the top 10 featurecombinations of the behavioral set
1000800
00900
1000200
0000
800900
100200
900200
400
MNIVARNOPSTDSKE
KUROFVN
OFMMDIS
ENTCONINDCORUNIIDM
200 400 600 800 1000 120000Occurrence probability ()
Figure 5 Occurrence analysis of the features in the behavioralregion
The top features based on the probability occurrence inFigure 5 are COR ENT DIS SKE STD VAR and MNIHowever the combination of these seven features does notresult in the best solution for classification Table 5 containsthe classification performance of the combination of featuresin the behavioral solution set In order to evaluate the per-formance of each feature combination various performancemeasures have been used such as precision sensitivity F-measure and accuracy Precision measures the fraction ofpositive instances from the group that the classifier predictedto be positive and recall measures the fraction of positiveexamples from the positive group of the actual class and [35]F-measure is the harmonic mean and accuracy is the pro-portion of true resultsThesemeasures can bemathematicallydefined as
Precision = TPTP + FP
Sensitivity = TPTP + FN
119865-Measure =2 (Sensitivity sdot Precision)Sensitivity + Precision
Accuracy = TPTP + FP + FN
(10)
where TP is correctly classified positive cases FP is incor-rectly classified negative cases and FN is incorrectly classifiedpositive cases For the performance measurement confusionmatrixes were created as described in Appendix and appliedinto (10) In the precision index number 1 combinationshows the highest performance in the behavioral solutionset while index number 7 combination shows the lowestIn the sensitivity index number 7 combination records thehighest results while index number 4 does the lowest In theF-measure and accuracy index number 2 combination showsthe highest record while index number 4 does the lowestBased on the confusion matrixes most of misclassificationoccurs in the classification of smoke smoke-reflection andother hot objects because during small fire texture patternsof these classes were diminished and the intensity was too low
8 Journal of Sensors
Table 5 Results of error case and performance at each feature combination in the behavioral solution set (Resu means resubstitution andCros refers to cross-validation)
Index Combination of features Error () Case Performance ()Resu Cros TP FP FN Precision Sensitivity 119865-measure Accuracy
1 MNI VAR ENT COR SKE 690 689 10049 402 324 9615 9688 9651 93262 MNI DIS COR SKE STD 668 670 10069 459 247 9564 9761 9662 93453 MNI ENT COR SKE STD 676 675 10061 447 267 9575 9741 9657 9337
4 MNI VAR DIS ENT CON CORSKE STD 698 690 9980 454 341 9565 9670 9617 9262
5 MNI VAR DIS ENT COR UNISKE STD 689 690 10037 453 285 9568 9724 9645 9315
6 MNI VAR DIS ENT COR IDMSKE STD 677 679 10047 461 267 9561 9741 9650 9324
7 MNI VAR DIS ENT IDM SKEKUR STD 696 694 10033 505 237 9521 9769 9643 9311
8 MNI VAR DIS ENT IND CORUNI SKE STD 695 699 10029 451 295 9570 9714 9641 9308
9 MNI VAR DIS ENT IND CORIDM SKE STD 688 690 10035 457 283 9564 9726 9644 9313
10 MNI VAR DIS ENT COR IDMSKE KUR STD 689 691 10036 478 261 9545 9747 9645 9314
PrecisionSensitivity
F-measureAccuracy
90919293949596979899
Perfo
rman
ce (
)
2 3 4 5 6 7 8 9 101Feature combination index
Figure 6 Results of performance at each feature combination in thebehavioral solution set
to distinguish The best solution was determined to be indexnumber 2 combination of MNI DIS COR SKE and STDwhich has the lowest of resubstitution and cross-validationerrors 668 and 670 respectively This combinationincludes all of the top features based on the probability occur-rence except ENT and VAR The four performance results ateach feature combination in the behavioral solution set areshown in Figure 6 where the highest results aremarked in redcircles and the lowest in green-dot circles Sensitivity appearshigher than precision at each feature combination becauseFPs are larger than FNs in the confusion matrix Particularlyindex number 7 has the biggest difference between FP and FNresulting in the highest sensitivity and lowest precision Thesummation of FP and FN in index number 4 is the highestin the behavioral solution set resulting in the lowest accuracywhile index number 2 has the lowest summation of FP andFN providing the highest accuracy
This study investigated a wide range of features fromlong-wavelength infrared camera images analyzed normaldistributions of fifteen features with respect to the classes ofsmoke fire and their thermal reflections and discovered thehighest performing feature combination by examining singlefeatures and multiple feature combinations As a result theproposed feature combination of MNI DIS COR SKE andSTD increases the performance compared with the previousstudy [1] which used MNI VAR ENT and IDM As shownin Figure 7 the errors are reduced by 286 and 268 resub-stitution and cross-validation errors and performances areincreased by 290 158 020 and 285 accuracy F-measure sensitivity and precision respectively
Figure 8 shows original visual and thermal images withthe robot at three different locations start point hallwayentrance and room entrance described in the experimentalfacility Each row relates to a series of images from the robotat three locations The first row contains visible images ofthe robot view As seen in the visible image at start pointfurther information regarding the hallway is limited due toshadowing of the light The image at hallway entrance showsa smoke layer in the upper portion of the hallway due to afire inside the roomThe image at the room entrance displaysa wood crib fire with sparks Because of soot and relativedifference in brightness the background is shown darker andthus limiting information on the background around the fire
Thermal infrared images are displayed in the second rowto show information that RGB camera cannot provide infire environments Unlike visual image at start point thatis obscured due to shadowing the presence of smoke andits thermal reflections on the ventilation hood can be obvi-ously perceived The red boxes on thermal images indicateobjects extracted through the adaptive object extraction withoptical flows and identification numbers In spite of dense
Journal of Sensors 9
9279
9741
9504
9055
954
938
9564
9761
9662
9345
668
670
Precision
Sensitivity
F-measure
Accuracy
Resubstitution error
Cross-validation error
The proposedThe previous
2000 4000 6000 8000 10000000()
Figure 7 Result comparison between the previous and the proposed studies
① Start point ② Hallway entrance ③ Room entrance
IR im
age
class
ifica
tion
resu
lts
Visib
le im
age
IR im
age
imag
e ana
lysis
Figure 8 Original visible and IR images with image analysis and classification results at different locations in the test setup of actual fires
smoke-filled and low visibility environments thermal imagescan generate the images of smoke and fire as well as back-ground information that is otherwise not visible throughvisual imaging
On the third row class labels and posterior probabilitiesof each candidate are displayed at the center of candidateROI as a result of Bayesian classification Using enhancedimage processing techniques the thermal images can bemore
10 Journal of Sensors
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
2
0
19
4820
64
Feature combination index number 2
Feature combination index number 3 Feature combination index number 4
Feature combination index number 5
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1462
0
0
0
1
1462
0
0
0
1
1457
0
0
0
1
1457
0
0
1365
193
197
0
0
1343
181
196
0
0
1325
205
181
0
0
1270
207
181
0
0
53
167
1929
0
0
145
166
1943
0
0
101
165
1942
7
0
49
4844
62
2
0
30
4817
63
0
0
53
165
1929
7
0
27
4830
61
MNI DIS COR SKE STD
MNI ENT COR SKE STD MNI VAR DIS ENT CON COR SKE STD
MNI VAR DIS ENT COR UNI SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bjPredicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Feature combination index number 1
140
159
0
0
0
0
0
00
0
0
0
0
0
1981
488
1280
193
144
1
1462
4838
62
25
2
MNI VAR ENT COR SKE
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Feature combination index number 6
0
0
0
0
488
0
0
0
1
14620
0
1339
206
185
0
0
93
161
1935
2
0
13
4823
67
MNI VAR DIS ENT COR IDM SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
F-re
flect
ion
Oth
er o
bj
Smok
e
Fire
S-re
flect
ion
Predicted class
Figure 9 Continued
Journal of Sensors 11
Feature combination index number 7 Feature combination index number 8
Feature combination index number 9 Feature combination index number 10
89
133
0
0
0
0
0
02
0
0
0
0
0
1888
488
0
0
0
0
488
1341
200
235
1
1459
4857
62
15
5
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1458
0
0
0
1
1462
0
0
0
1
1462
0
0
1324
204
181
0
0
1334
208
181
0
0
1334
205
204
0
0
99
174
1943
0
0
91
150
1921
2
0
18
4810
65
6
0
20
4835
62
0
0
93
172
1941
2
0
22
4812
63
MNI VAR DIS ENT COR IDM SKE KUR STD
MNI VAR DIS ENT IDM SKE KUR STD
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bj
Predicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
MNI VAR DIS ENT IND COR IDM SKE STD
MNI VAR DIS ENT IND COR UNI SKE STD
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
Figure 9 Results of confusion matrix at each feature combination
refined and clearer than the thermal images on the secondrow Smoke fire and their thermal reflections are identifiedand marked in red or orange ellipses
6 Conclusion
The appropriate combination of features was investigated toaccurately classify fire smoke and their thermal reflectionsusing thermal images Gray-scale 14-bit images from a singleinfrared camera were used to extract motion and texture fea-tures by applying a clustering-based autothresholding tech-nique Bayesian classification is performed to probabilisti-cally identify multiple classes during real-time implemen-tation To find the best combination of features a multi-objective genetic algorithm optimization was implementedusing resubstitution and cross-validation errors as objectivefunctions Large-scale fire tests with different fire sources
were conducted to create a range of temperature and smokeconditions to evaluate the feature combinations
Fifteenmotion and texture features were analyzed and theprobability density functions of the features were computedby the maximum likelihood estimation The combinationof multiple features was determined to more accuratelyclassify fire smoke and thermal reflections compared witha single feature In the behavioral solution set where featurecombinations produce less than 7 resubstitution and cross-validation errors COR ENT DIS SKE STD VAR andMNIhad 800 or more occurrence while other features had400 or less occurrence The feature combination of MNIDIS COR SKE and STD produced the highest performancein the classification resulting in 668 and 670 resubstitu-tion and cross-validation errors and 9564 9761 9662and 9345 precision sensitivity F-measure and accuracyrespectively
12 Journal of Sensors
In the near future the classification of fire smoke andtheir thermal reflections will be evaluated on any classifiersand features to increase performanceThe convolution neuralnetwork of deep learning which has recently shown highperformance could be explored as a classifier also model-based image features such as discrete wavelet transform willbe further studied
Appendix
See Figure 9
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this manuscript
Acknowledgments
This work was sponsored by the Office of Naval ResearchGrant no N00014-11-1-0074 scientific office Dr ThomasMcKenna in USA Hwarang-dae Research Institute in Seouland Agency for Defense Development in Daejeon SouthKorea The authors would like to thank Mr Joseph Starr andMr JoshMcNeil for assisting in performing the fire testsTheauthors would also like to thank Rosana K Lee for helpingand supporting this research
References
[1] J-H Kim and B Y Lattimer ldquoReal-time probabilistic classifi-cation of fire and smoke using thermal imagery for intelligentfirefighting robotrdquo Fire Safety Journal vol 72 pp 40ndash49 2015
[2] J W Starr and B Y Lattimer ldquoEvaluation of navigation sensorsin fire smoke environmentsrdquo Fire Technology vol 50 no 6 pp1459ndash1481 2014
[3] J-H Kim B Keller and B Y Lattimer ldquoSensor fusion basedseek-and-find fire algorithm for intelligent firefighting robotrdquoin Proceedings of the IEEEASME International Conference onAdvanced Intelligent Mechatronics (AIM rsquo13) pp 1482ndash1486IEEE Wollongong Australia July 2013
[4] J G McNeil J Starr and B Y Lattimer ldquoAutonomous fire sup-pression using multispectral sensorsrdquo in Proceedings of theIEEEASME International Conference on Advanced IntelligentMechatronics Mechatronics for HumanWellbeing (AIM rsquo13) pp1504ndash1509 Wollongong Austrailia July 2013
[5] J-H Kim J W Starr and B Y Lattimer ldquoFirefighting robotstereo infrared vision and radar sensor fusion for imagingthrough smokerdquo Fire Technology vol 51 no 4 pp 823ndash8452015
[6] RC Luo andK L Su ldquoAutonomous fire-detection systemusingadaptive sensory fusion for intelligent security robotrdquo IEEEASME Transactions on Mechatronics vol 12 no 3 pp 274ndash2812007
[7] M A Jackson and I Robins ldquoGas sensing for fire detectionmeasurements of CO CO
2
H2
O2
and smoke density in Euro-pean standard fire testsrdquo Fire Safety Journal vol 22 no 2 pp181ndash205 1994
[8] B U Toreyin R G Cinbis Y Dedeoglu and A E Cetin ldquoFiredetection in infrared video using wavelet analysisrdquo OpticalEngineering vol 46 no 6 Article ID 067204 2007
[9] B U Toreyin Y Dedeoglu andA E Cetin ldquoWavelet based real-time smoke detection in videordquo in Proceedings of the 13th Euro-pean Signal Processing Conference pp 4ndash8 Antalya TurkeySeptember 2005
[10] T Celik H Demirel H Ozkaramanli and M Uyguroglu ldquoFiredetection using statistical color model in video sequencesrdquoJournal of Visual Communication and Image Representation vol18 no 2 pp 176ndash185 2007
[11] L Merino F Caballero J R Martınez-de-Dios I Maza and AOllero ldquoAn unmanned aircraft system for automatic forest firemonitoring and measurementrdquo Journal of Intelligent amp RoboticSystems vol 65 no 1 pp 533ndash548 2012
[12] Y Wang T W Chua R Chang and N T Pham ldquoReal-timesmoke detection using texture and color featuresrdquo in Proceed-ings of the 21st International Conference on Pattern Recognition(ICPR rsquo12) pp 1727ndash1730 Tsukuba Japan November 2012
[13] G Marbach M Loepfe and T Brupbacher ldquoAn image process-ing technique for fire detection in video imagesrdquo Fire SafetyJournal vol 41 no 4 pp 285ndash289 2006
[14] M I Chacon-Murguia and F J Perez-Vargas ldquoThermal videoanalysis for fire detection using shape regularity and intensitysaturation featuresrdquo in Pattern Recognition J F Martınez-Trinidad J A Carrasco-Ochoa C B-Y Brants and E RHancock Eds vol 6718 of Lecture Notes in Computer Sciencepp 118ndash126 Springer Berlin Germany 2011
[15] W Phillips III M Shah and N D V Lobo ldquoFlame recognitionin videordquo Pattern Recognition Letters vol 23 no 1ndash3 pp 319ndash327 2002
[16] D Han and B Lee ldquoDevelopment of early tunnel fire detectionalgorithm using the image processingrdquo in Advances in VisualComputing pp 39ndash48 Springer Berlin Germany 2006
[17] Y Chunyu F Jun W Jinjun and Z Yongming ldquoVideo firesmoke detection using motion and color featuresrdquo Fire Technol-ogy vol 46 no 3 pp 651ndash663 2010
[18] F Yuan ldquoVideo-based smoke detection with histogram se-quence of LBP and LBPV pyramidsrdquo Fire Safety Journal vol 46no 3 pp 132ndash139 2011
[19] F Lafarge X Descombes and J Zerubia ldquoTextural kernel forSVM classification in remote sensing application to forest firedetection andUrban area extractionrdquo in Proceedings of the IEEEInternational Conference on Image Processing (ICIP rsquo05) pp1096ndash1099 September 2005
[20] F Amon and A Ducharme ldquoImage frequency analysis for test-ing of fire service thermal imaging camerasrdquo Fire Technologyvol 45 no 3 pp 313ndash322 2009
[21] F Amon V Benetis J Kim and A Hamins ldquoDevelopmentof a performance evaluation facility for fire fighting thermalimagersrdquo in Defense and Security pp 244ndash252 2004
[22] F D Maxwell ldquoA portable IR system for observing fire thrusmokerdquo Fire Technology vol 7 no 4 pp 321ndash331 1971
[23] A Barducci D Guzzi P Marcoionni and I Pippi ldquoInfrareddetection of active fires and burnt areas theory and observa-tionsrdquo Infrared Physics amp Technology vol 43 no 3ndash5 pp 119ndash125 2002
[24] C Wang and S Qin ldquoAdaptive detection method of infraredsmall target based on target-background separation via robustprincipal component analysisrdquo Infrared Physics amp Technologyvol 69 pp 123ndash135 2015
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
6 Journal of Sensors
0 5 10VAR
Prob
abili
ty d
ensit
y
SmokeS-reflection
FireF-reflection
times10minus4
times10minus4
times104
times10minus7
times106
times10minus3
2468
1012
Prob
abili
ty d
ensit
y
minus10000 minus8000 minus6000 minus4000minus12000MNI
0
5
10
Prob
abili
ty d
ensit
y
0
05
1
Prob
abili
ty d
ensit
y
5000 100000OFVN
0
20
40
Prob
abili
ty d
ensit
y
01 02 03 040DIS
500 1000 1500 20000OFVMM
0
0005
001
Prob
abili
ty d
ensit
y0
5
Prob
abili
ty d
ensit
y
2 4 60NOP
50100150200250
Prob
abili
ty d
ensit
y
09 095 1085IND
1 15 2 2505CON
2
4
6
Prob
abili
ty d
ensit
y
02 04 06 08 10ENT
0
5
Prob
abili
ty d
ensit
y
2468
1012
Prob
abili
ty d
ensit
y
06 08 104COR
minus2 0 2 4 6minus4SKE
0
05
1
Prob
abili
ty d
ensit
y
01
02
03
Prob
abili
ty d
ensit
y
5 10 15 200KUT
1000 2000 3000 40000STD
0
2
4
08 10604IDM
510152025
Prob
abili
ty d
ensit
y
02 04 06 08 10UNI
0
10
20
Prob
abili
ty d
ensit
y
SmokeS-reflection
FireF-reflection
SmokeS-reflection
FireF-reflection
times10minus3
Figure 3 Probability density distributions of each feature
statistical texture features show42sim45error which is higherthan the other second-order features
52 Multiple Feature Combination Performance The errorresults in Table 4 demonstrate that a single feature cannotaccurately classify fire smoke and thermal reflections Thuspossible combination ofmultiple features was considered andanalyzed to find the best combination of the featuresThe totalnumber of all possible combinations that have two or morefeatures is
119873total =119898
sum
119911=2
(
119898
119911
) (9)
where 119898 refers to the total number of features (ie 119898 = 15)and 119911 is the number of features in the combination Basedon all possible combination the multiobjective genetic algo-rithm optimization [40] in the global optimization toolbox ofMATLAB was used to find the best combination of featuresthat has the highest performance in the classification Theobjective functions in the optimization resubstitution and119896-fold cross-validation errors [39] were used to measurehow accurately different feature combinations perform in theclassification
Figure 4 contains a plot of the error associated with themost promising feature combinations The behavioral solu-tion set is defined as feature combinations with less than 7
Journal of Sensors 7
Table 4 Performance of each feature
Resubstitution error () Cross-validation error ()MNI 237 238VAR 243 243NOP 721 720STD 232 232SKE 527 528KUR 506 506OFVN 395 400OFVMM 586 586DIS 290 290ENT 446 446CON 285 285IND 301 301COR 412 411UNI 345 345IDM 377 377
Behavioral region
Behavioral solution setGeneral set
66 67 68 69 7 71 72 7365Resubstitute error ()
65
66
67
68
69
7
71
72
73
Cros
s-va
lidat
ion
erro
r (
)
Figure 4Multiobjective optimization result showing the general setand behavioral solution set (colored region)
error for both objective functions while the general set refersto all other possible feature combinations The behavioralsolution set contains 00061 of all possible feature combi-nations
The occurrence probability of features in the behavioralsolution set is illustrated in Figure 5 In the behavioralsolution set the first-order statistic texture features MNI andSKE always exist while OFVN NOP and OFVMM featuresdo not Both the first-order statistical texture features STDand VAR and the second-order statistical texture featuresCOR ENT andDIS showahigher occurrence comparedwiththe other first- and second-order texture features while KURIDMUNI IND andCON show lower occurrence Note thatdue to the robotrsquos dynamical motion motion features werenot successful and even not included in the top 10 featurecombinations of the behavioral set
1000800
00900
1000200
0000
800900
100200
900200
400
MNIVARNOPSTDSKE
KUROFVN
OFMMDIS
ENTCONINDCORUNIIDM
200 400 600 800 1000 120000Occurrence probability ()
Figure 5 Occurrence analysis of the features in the behavioralregion
The top features based on the probability occurrence inFigure 5 are COR ENT DIS SKE STD VAR and MNIHowever the combination of these seven features does notresult in the best solution for classification Table 5 containsthe classification performance of the combination of featuresin the behavioral solution set In order to evaluate the per-formance of each feature combination various performancemeasures have been used such as precision sensitivity F-measure and accuracy Precision measures the fraction ofpositive instances from the group that the classifier predictedto be positive and recall measures the fraction of positiveexamples from the positive group of the actual class and [35]F-measure is the harmonic mean and accuracy is the pro-portion of true resultsThesemeasures can bemathematicallydefined as
Precision = TPTP + FP
Sensitivity = TPTP + FN
119865-Measure =2 (Sensitivity sdot Precision)Sensitivity + Precision
Accuracy = TPTP + FP + FN
(10)
where TP is correctly classified positive cases FP is incor-rectly classified negative cases and FN is incorrectly classifiedpositive cases For the performance measurement confusionmatrixes were created as described in Appendix and appliedinto (10) In the precision index number 1 combinationshows the highest performance in the behavioral solutionset while index number 7 combination shows the lowestIn the sensitivity index number 7 combination records thehighest results while index number 4 does the lowest In theF-measure and accuracy index number 2 combination showsthe highest record while index number 4 does the lowestBased on the confusion matrixes most of misclassificationoccurs in the classification of smoke smoke-reflection andother hot objects because during small fire texture patternsof these classes were diminished and the intensity was too low
8 Journal of Sensors
Table 5 Results of error case and performance at each feature combination in the behavioral solution set (Resu means resubstitution andCros refers to cross-validation)
Index Combination of features Error () Case Performance ()Resu Cros TP FP FN Precision Sensitivity 119865-measure Accuracy
1 MNI VAR ENT COR SKE 690 689 10049 402 324 9615 9688 9651 93262 MNI DIS COR SKE STD 668 670 10069 459 247 9564 9761 9662 93453 MNI ENT COR SKE STD 676 675 10061 447 267 9575 9741 9657 9337
4 MNI VAR DIS ENT CON CORSKE STD 698 690 9980 454 341 9565 9670 9617 9262
5 MNI VAR DIS ENT COR UNISKE STD 689 690 10037 453 285 9568 9724 9645 9315
6 MNI VAR DIS ENT COR IDMSKE STD 677 679 10047 461 267 9561 9741 9650 9324
7 MNI VAR DIS ENT IDM SKEKUR STD 696 694 10033 505 237 9521 9769 9643 9311
8 MNI VAR DIS ENT IND CORUNI SKE STD 695 699 10029 451 295 9570 9714 9641 9308
9 MNI VAR DIS ENT IND CORIDM SKE STD 688 690 10035 457 283 9564 9726 9644 9313
10 MNI VAR DIS ENT COR IDMSKE KUR STD 689 691 10036 478 261 9545 9747 9645 9314
PrecisionSensitivity
F-measureAccuracy
90919293949596979899
Perfo
rman
ce (
)
2 3 4 5 6 7 8 9 101Feature combination index
Figure 6 Results of performance at each feature combination in thebehavioral solution set
to distinguish The best solution was determined to be indexnumber 2 combination of MNI DIS COR SKE and STDwhich has the lowest of resubstitution and cross-validationerrors 668 and 670 respectively This combinationincludes all of the top features based on the probability occur-rence except ENT and VAR The four performance results ateach feature combination in the behavioral solution set areshown in Figure 6 where the highest results aremarked in redcircles and the lowest in green-dot circles Sensitivity appearshigher than precision at each feature combination becauseFPs are larger than FNs in the confusion matrix Particularlyindex number 7 has the biggest difference between FP and FNresulting in the highest sensitivity and lowest precision Thesummation of FP and FN in index number 4 is the highestin the behavioral solution set resulting in the lowest accuracywhile index number 2 has the lowest summation of FP andFN providing the highest accuracy
This study investigated a wide range of features fromlong-wavelength infrared camera images analyzed normaldistributions of fifteen features with respect to the classes ofsmoke fire and their thermal reflections and discovered thehighest performing feature combination by examining singlefeatures and multiple feature combinations As a result theproposed feature combination of MNI DIS COR SKE andSTD increases the performance compared with the previousstudy [1] which used MNI VAR ENT and IDM As shownin Figure 7 the errors are reduced by 286 and 268 resub-stitution and cross-validation errors and performances areincreased by 290 158 020 and 285 accuracy F-measure sensitivity and precision respectively
Figure 8 shows original visual and thermal images withthe robot at three different locations start point hallwayentrance and room entrance described in the experimentalfacility Each row relates to a series of images from the robotat three locations The first row contains visible images ofthe robot view As seen in the visible image at start pointfurther information regarding the hallway is limited due toshadowing of the light The image at hallway entrance showsa smoke layer in the upper portion of the hallway due to afire inside the roomThe image at the room entrance displaysa wood crib fire with sparks Because of soot and relativedifference in brightness the background is shown darker andthus limiting information on the background around the fire
Thermal infrared images are displayed in the second rowto show information that RGB camera cannot provide infire environments Unlike visual image at start point thatis obscured due to shadowing the presence of smoke andits thermal reflections on the ventilation hood can be obvi-ously perceived The red boxes on thermal images indicateobjects extracted through the adaptive object extraction withoptical flows and identification numbers In spite of dense
Journal of Sensors 9
9279
9741
9504
9055
954
938
9564
9761
9662
9345
668
670
Precision
Sensitivity
F-measure
Accuracy
Resubstitution error
Cross-validation error
The proposedThe previous
2000 4000 6000 8000 10000000()
Figure 7 Result comparison between the previous and the proposed studies
① Start point ② Hallway entrance ③ Room entrance
IR im
age
class
ifica
tion
resu
lts
Visib
le im
age
IR im
age
imag
e ana
lysis
Figure 8 Original visible and IR images with image analysis and classification results at different locations in the test setup of actual fires
smoke-filled and low visibility environments thermal imagescan generate the images of smoke and fire as well as back-ground information that is otherwise not visible throughvisual imaging
On the third row class labels and posterior probabilitiesof each candidate are displayed at the center of candidateROI as a result of Bayesian classification Using enhancedimage processing techniques the thermal images can bemore
10 Journal of Sensors
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
2
0
19
4820
64
Feature combination index number 2
Feature combination index number 3 Feature combination index number 4
Feature combination index number 5
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1462
0
0
0
1
1462
0
0
0
1
1457
0
0
0
1
1457
0
0
1365
193
197
0
0
1343
181
196
0
0
1325
205
181
0
0
1270
207
181
0
0
53
167
1929
0
0
145
166
1943
0
0
101
165
1942
7
0
49
4844
62
2
0
30
4817
63
0
0
53
165
1929
7
0
27
4830
61
MNI DIS COR SKE STD
MNI ENT COR SKE STD MNI VAR DIS ENT CON COR SKE STD
MNI VAR DIS ENT COR UNI SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bjPredicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Feature combination index number 1
140
159
0
0
0
0
0
00
0
0
0
0
0
1981
488
1280
193
144
1
1462
4838
62
25
2
MNI VAR ENT COR SKE
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Feature combination index number 6
0
0
0
0
488
0
0
0
1
14620
0
1339
206
185
0
0
93
161
1935
2
0
13
4823
67
MNI VAR DIS ENT COR IDM SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
F-re
flect
ion
Oth
er o
bj
Smok
e
Fire
S-re
flect
ion
Predicted class
Figure 9 Continued
Journal of Sensors 11
Feature combination index number 7 Feature combination index number 8
Feature combination index number 9 Feature combination index number 10
89
133
0
0
0
0
0
02
0
0
0
0
0
1888
488
0
0
0
0
488
1341
200
235
1
1459
4857
62
15
5
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1458
0
0
0
1
1462
0
0
0
1
1462
0
0
1324
204
181
0
0
1334
208
181
0
0
1334
205
204
0
0
99
174
1943
0
0
91
150
1921
2
0
18
4810
65
6
0
20
4835
62
0
0
93
172
1941
2
0
22
4812
63
MNI VAR DIS ENT COR IDM SKE KUR STD
MNI VAR DIS ENT IDM SKE KUR STD
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bj
Predicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
MNI VAR DIS ENT IND COR IDM SKE STD
MNI VAR DIS ENT IND COR UNI SKE STD
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
Figure 9 Results of confusion matrix at each feature combination
refined and clearer than the thermal images on the secondrow Smoke fire and their thermal reflections are identifiedand marked in red or orange ellipses
6 Conclusion
The appropriate combination of features was investigated toaccurately classify fire smoke and their thermal reflectionsusing thermal images Gray-scale 14-bit images from a singleinfrared camera were used to extract motion and texture fea-tures by applying a clustering-based autothresholding tech-nique Bayesian classification is performed to probabilisti-cally identify multiple classes during real-time implemen-tation To find the best combination of features a multi-objective genetic algorithm optimization was implementedusing resubstitution and cross-validation errors as objectivefunctions Large-scale fire tests with different fire sources
were conducted to create a range of temperature and smokeconditions to evaluate the feature combinations
Fifteenmotion and texture features were analyzed and theprobability density functions of the features were computedby the maximum likelihood estimation The combinationof multiple features was determined to more accuratelyclassify fire smoke and thermal reflections compared witha single feature In the behavioral solution set where featurecombinations produce less than 7 resubstitution and cross-validation errors COR ENT DIS SKE STD VAR andMNIhad 800 or more occurrence while other features had400 or less occurrence The feature combination of MNIDIS COR SKE and STD produced the highest performancein the classification resulting in 668 and 670 resubstitu-tion and cross-validation errors and 9564 9761 9662and 9345 precision sensitivity F-measure and accuracyrespectively
12 Journal of Sensors
In the near future the classification of fire smoke andtheir thermal reflections will be evaluated on any classifiersand features to increase performanceThe convolution neuralnetwork of deep learning which has recently shown highperformance could be explored as a classifier also model-based image features such as discrete wavelet transform willbe further studied
Appendix
See Figure 9
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this manuscript
Acknowledgments
This work was sponsored by the Office of Naval ResearchGrant no N00014-11-1-0074 scientific office Dr ThomasMcKenna in USA Hwarang-dae Research Institute in Seouland Agency for Defense Development in Daejeon SouthKorea The authors would like to thank Mr Joseph Starr andMr JoshMcNeil for assisting in performing the fire testsTheauthors would also like to thank Rosana K Lee for helpingand supporting this research
References
[1] J-H Kim and B Y Lattimer ldquoReal-time probabilistic classifi-cation of fire and smoke using thermal imagery for intelligentfirefighting robotrdquo Fire Safety Journal vol 72 pp 40ndash49 2015
[2] J W Starr and B Y Lattimer ldquoEvaluation of navigation sensorsin fire smoke environmentsrdquo Fire Technology vol 50 no 6 pp1459ndash1481 2014
[3] J-H Kim B Keller and B Y Lattimer ldquoSensor fusion basedseek-and-find fire algorithm for intelligent firefighting robotrdquoin Proceedings of the IEEEASME International Conference onAdvanced Intelligent Mechatronics (AIM rsquo13) pp 1482ndash1486IEEE Wollongong Australia July 2013
[4] J G McNeil J Starr and B Y Lattimer ldquoAutonomous fire sup-pression using multispectral sensorsrdquo in Proceedings of theIEEEASME International Conference on Advanced IntelligentMechatronics Mechatronics for HumanWellbeing (AIM rsquo13) pp1504ndash1509 Wollongong Austrailia July 2013
[5] J-H Kim J W Starr and B Y Lattimer ldquoFirefighting robotstereo infrared vision and radar sensor fusion for imagingthrough smokerdquo Fire Technology vol 51 no 4 pp 823ndash8452015
[6] RC Luo andK L Su ldquoAutonomous fire-detection systemusingadaptive sensory fusion for intelligent security robotrdquo IEEEASME Transactions on Mechatronics vol 12 no 3 pp 274ndash2812007
[7] M A Jackson and I Robins ldquoGas sensing for fire detectionmeasurements of CO CO
2
H2
O2
and smoke density in Euro-pean standard fire testsrdquo Fire Safety Journal vol 22 no 2 pp181ndash205 1994
[8] B U Toreyin R G Cinbis Y Dedeoglu and A E Cetin ldquoFiredetection in infrared video using wavelet analysisrdquo OpticalEngineering vol 46 no 6 Article ID 067204 2007
[9] B U Toreyin Y Dedeoglu andA E Cetin ldquoWavelet based real-time smoke detection in videordquo in Proceedings of the 13th Euro-pean Signal Processing Conference pp 4ndash8 Antalya TurkeySeptember 2005
[10] T Celik H Demirel H Ozkaramanli and M Uyguroglu ldquoFiredetection using statistical color model in video sequencesrdquoJournal of Visual Communication and Image Representation vol18 no 2 pp 176ndash185 2007
[11] L Merino F Caballero J R Martınez-de-Dios I Maza and AOllero ldquoAn unmanned aircraft system for automatic forest firemonitoring and measurementrdquo Journal of Intelligent amp RoboticSystems vol 65 no 1 pp 533ndash548 2012
[12] Y Wang T W Chua R Chang and N T Pham ldquoReal-timesmoke detection using texture and color featuresrdquo in Proceed-ings of the 21st International Conference on Pattern Recognition(ICPR rsquo12) pp 1727ndash1730 Tsukuba Japan November 2012
[13] G Marbach M Loepfe and T Brupbacher ldquoAn image process-ing technique for fire detection in video imagesrdquo Fire SafetyJournal vol 41 no 4 pp 285ndash289 2006
[14] M I Chacon-Murguia and F J Perez-Vargas ldquoThermal videoanalysis for fire detection using shape regularity and intensitysaturation featuresrdquo in Pattern Recognition J F Martınez-Trinidad J A Carrasco-Ochoa C B-Y Brants and E RHancock Eds vol 6718 of Lecture Notes in Computer Sciencepp 118ndash126 Springer Berlin Germany 2011
[15] W Phillips III M Shah and N D V Lobo ldquoFlame recognitionin videordquo Pattern Recognition Letters vol 23 no 1ndash3 pp 319ndash327 2002
[16] D Han and B Lee ldquoDevelopment of early tunnel fire detectionalgorithm using the image processingrdquo in Advances in VisualComputing pp 39ndash48 Springer Berlin Germany 2006
[17] Y Chunyu F Jun W Jinjun and Z Yongming ldquoVideo firesmoke detection using motion and color featuresrdquo Fire Technol-ogy vol 46 no 3 pp 651ndash663 2010
[18] F Yuan ldquoVideo-based smoke detection with histogram se-quence of LBP and LBPV pyramidsrdquo Fire Safety Journal vol 46no 3 pp 132ndash139 2011
[19] F Lafarge X Descombes and J Zerubia ldquoTextural kernel forSVM classification in remote sensing application to forest firedetection andUrban area extractionrdquo in Proceedings of the IEEEInternational Conference on Image Processing (ICIP rsquo05) pp1096ndash1099 September 2005
[20] F Amon and A Ducharme ldquoImage frequency analysis for test-ing of fire service thermal imaging camerasrdquo Fire Technologyvol 45 no 3 pp 313ndash322 2009
[21] F Amon V Benetis J Kim and A Hamins ldquoDevelopmentof a performance evaluation facility for fire fighting thermalimagersrdquo in Defense and Security pp 244ndash252 2004
[22] F D Maxwell ldquoA portable IR system for observing fire thrusmokerdquo Fire Technology vol 7 no 4 pp 321ndash331 1971
[23] A Barducci D Guzzi P Marcoionni and I Pippi ldquoInfrareddetection of active fires and burnt areas theory and observa-tionsrdquo Infrared Physics amp Technology vol 43 no 3ndash5 pp 119ndash125 2002
[24] C Wang and S Qin ldquoAdaptive detection method of infraredsmall target based on target-background separation via robustprincipal component analysisrdquo Infrared Physics amp Technologyvol 69 pp 123ndash135 2015
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
Journal of Sensors 7
Table 4 Performance of each feature
Resubstitution error () Cross-validation error ()MNI 237 238VAR 243 243NOP 721 720STD 232 232SKE 527 528KUR 506 506OFVN 395 400OFVMM 586 586DIS 290 290ENT 446 446CON 285 285IND 301 301COR 412 411UNI 345 345IDM 377 377
Behavioral region
Behavioral solution setGeneral set
66 67 68 69 7 71 72 7365Resubstitute error ()
65
66
67
68
69
7
71
72
73
Cros
s-va
lidat
ion
erro
r (
)
Figure 4Multiobjective optimization result showing the general setand behavioral solution set (colored region)
error for both objective functions while the general set refersto all other possible feature combinations The behavioralsolution set contains 00061 of all possible feature combi-nations
The occurrence probability of features in the behavioralsolution set is illustrated in Figure 5 In the behavioralsolution set the first-order statistic texture features MNI andSKE always exist while OFVN NOP and OFVMM featuresdo not Both the first-order statistical texture features STDand VAR and the second-order statistical texture featuresCOR ENT andDIS showahigher occurrence comparedwiththe other first- and second-order texture features while KURIDMUNI IND andCON show lower occurrence Note thatdue to the robotrsquos dynamical motion motion features werenot successful and even not included in the top 10 featurecombinations of the behavioral set
1000800
00900
1000200
0000
800900
100200
900200
400
MNIVARNOPSTDSKE
KUROFVN
OFMMDIS
ENTCONINDCORUNIIDM
200 400 600 800 1000 120000Occurrence probability ()
Figure 5 Occurrence analysis of the features in the behavioralregion
The top features based on the probability occurrence inFigure 5 are COR ENT DIS SKE STD VAR and MNIHowever the combination of these seven features does notresult in the best solution for classification Table 5 containsthe classification performance of the combination of featuresin the behavioral solution set In order to evaluate the per-formance of each feature combination various performancemeasures have been used such as precision sensitivity F-measure and accuracy Precision measures the fraction ofpositive instances from the group that the classifier predictedto be positive and recall measures the fraction of positiveexamples from the positive group of the actual class and [35]F-measure is the harmonic mean and accuracy is the pro-portion of true resultsThesemeasures can bemathematicallydefined as
Precision = TPTP + FP
Sensitivity = TPTP + FN
119865-Measure =2 (Sensitivity sdot Precision)Sensitivity + Precision
Accuracy = TPTP + FP + FN
(10)
where TP is correctly classified positive cases FP is incor-rectly classified negative cases and FN is incorrectly classifiedpositive cases For the performance measurement confusionmatrixes were created as described in Appendix and appliedinto (10) In the precision index number 1 combinationshows the highest performance in the behavioral solutionset while index number 7 combination shows the lowestIn the sensitivity index number 7 combination records thehighest results while index number 4 does the lowest In theF-measure and accuracy index number 2 combination showsthe highest record while index number 4 does the lowestBased on the confusion matrixes most of misclassificationoccurs in the classification of smoke smoke-reflection andother hot objects because during small fire texture patternsof these classes were diminished and the intensity was too low
8 Journal of Sensors
Table 5 Results of error case and performance at each feature combination in the behavioral solution set (Resu means resubstitution andCros refers to cross-validation)
Index Combination of features Error () Case Performance ()Resu Cros TP FP FN Precision Sensitivity 119865-measure Accuracy
1 MNI VAR ENT COR SKE 690 689 10049 402 324 9615 9688 9651 93262 MNI DIS COR SKE STD 668 670 10069 459 247 9564 9761 9662 93453 MNI ENT COR SKE STD 676 675 10061 447 267 9575 9741 9657 9337
4 MNI VAR DIS ENT CON CORSKE STD 698 690 9980 454 341 9565 9670 9617 9262
5 MNI VAR DIS ENT COR UNISKE STD 689 690 10037 453 285 9568 9724 9645 9315
6 MNI VAR DIS ENT COR IDMSKE STD 677 679 10047 461 267 9561 9741 9650 9324
7 MNI VAR DIS ENT IDM SKEKUR STD 696 694 10033 505 237 9521 9769 9643 9311
8 MNI VAR DIS ENT IND CORUNI SKE STD 695 699 10029 451 295 9570 9714 9641 9308
9 MNI VAR DIS ENT IND CORIDM SKE STD 688 690 10035 457 283 9564 9726 9644 9313
10 MNI VAR DIS ENT COR IDMSKE KUR STD 689 691 10036 478 261 9545 9747 9645 9314
PrecisionSensitivity
F-measureAccuracy
90919293949596979899
Perfo
rman
ce (
)
2 3 4 5 6 7 8 9 101Feature combination index
Figure 6 Results of performance at each feature combination in thebehavioral solution set
to distinguish The best solution was determined to be indexnumber 2 combination of MNI DIS COR SKE and STDwhich has the lowest of resubstitution and cross-validationerrors 668 and 670 respectively This combinationincludes all of the top features based on the probability occur-rence except ENT and VAR The four performance results ateach feature combination in the behavioral solution set areshown in Figure 6 where the highest results aremarked in redcircles and the lowest in green-dot circles Sensitivity appearshigher than precision at each feature combination becauseFPs are larger than FNs in the confusion matrix Particularlyindex number 7 has the biggest difference between FP and FNresulting in the highest sensitivity and lowest precision Thesummation of FP and FN in index number 4 is the highestin the behavioral solution set resulting in the lowest accuracywhile index number 2 has the lowest summation of FP andFN providing the highest accuracy
This study investigated a wide range of features fromlong-wavelength infrared camera images analyzed normaldistributions of fifteen features with respect to the classes ofsmoke fire and their thermal reflections and discovered thehighest performing feature combination by examining singlefeatures and multiple feature combinations As a result theproposed feature combination of MNI DIS COR SKE andSTD increases the performance compared with the previousstudy [1] which used MNI VAR ENT and IDM As shownin Figure 7 the errors are reduced by 286 and 268 resub-stitution and cross-validation errors and performances areincreased by 290 158 020 and 285 accuracy F-measure sensitivity and precision respectively
Figure 8 shows original visual and thermal images withthe robot at three different locations start point hallwayentrance and room entrance described in the experimentalfacility Each row relates to a series of images from the robotat three locations The first row contains visible images ofthe robot view As seen in the visible image at start pointfurther information regarding the hallway is limited due toshadowing of the light The image at hallway entrance showsa smoke layer in the upper portion of the hallway due to afire inside the roomThe image at the room entrance displaysa wood crib fire with sparks Because of soot and relativedifference in brightness the background is shown darker andthus limiting information on the background around the fire
Thermal infrared images are displayed in the second rowto show information that RGB camera cannot provide infire environments Unlike visual image at start point thatis obscured due to shadowing the presence of smoke andits thermal reflections on the ventilation hood can be obvi-ously perceived The red boxes on thermal images indicateobjects extracted through the adaptive object extraction withoptical flows and identification numbers In spite of dense
Journal of Sensors 9
9279
9741
9504
9055
954
938
9564
9761
9662
9345
668
670
Precision
Sensitivity
F-measure
Accuracy
Resubstitution error
Cross-validation error
The proposedThe previous
2000 4000 6000 8000 10000000()
Figure 7 Result comparison between the previous and the proposed studies
① Start point ② Hallway entrance ③ Room entrance
IR im
age
class
ifica
tion
resu
lts
Visib
le im
age
IR im
age
imag
e ana
lysis
Figure 8 Original visible and IR images with image analysis and classification results at different locations in the test setup of actual fires
smoke-filled and low visibility environments thermal imagescan generate the images of smoke and fire as well as back-ground information that is otherwise not visible throughvisual imaging
On the third row class labels and posterior probabilitiesof each candidate are displayed at the center of candidateROI as a result of Bayesian classification Using enhancedimage processing techniques the thermal images can bemore
10 Journal of Sensors
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
2
0
19
4820
64
Feature combination index number 2
Feature combination index number 3 Feature combination index number 4
Feature combination index number 5
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1462
0
0
0
1
1462
0
0
0
1
1457
0
0
0
1
1457
0
0
1365
193
197
0
0
1343
181
196
0
0
1325
205
181
0
0
1270
207
181
0
0
53
167
1929
0
0
145
166
1943
0
0
101
165
1942
7
0
49
4844
62
2
0
30
4817
63
0
0
53
165
1929
7
0
27
4830
61
MNI DIS COR SKE STD
MNI ENT COR SKE STD MNI VAR DIS ENT CON COR SKE STD
MNI VAR DIS ENT COR UNI SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bjPredicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Feature combination index number 1
140
159
0
0
0
0
0
00
0
0
0
0
0
1981
488
1280
193
144
1
1462
4838
62
25
2
MNI VAR ENT COR SKE
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Feature combination index number 6
0
0
0
0
488
0
0
0
1
14620
0
1339
206
185
0
0
93
161
1935
2
0
13
4823
67
MNI VAR DIS ENT COR IDM SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
F-re
flect
ion
Oth
er o
bj
Smok
e
Fire
S-re
flect
ion
Predicted class
Figure 9 Continued
Journal of Sensors 11
Feature combination index number 7 Feature combination index number 8
Feature combination index number 9 Feature combination index number 10
89
133
0
0
0
0
0
02
0
0
0
0
0
1888
488
0
0
0
0
488
1341
200
235
1
1459
4857
62
15
5
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1458
0
0
0
1
1462
0
0
0
1
1462
0
0
1324
204
181
0
0
1334
208
181
0
0
1334
205
204
0
0
99
174
1943
0
0
91
150
1921
2
0
18
4810
65
6
0
20
4835
62
0
0
93
172
1941
2
0
22
4812
63
MNI VAR DIS ENT COR IDM SKE KUR STD
MNI VAR DIS ENT IDM SKE KUR STD
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bj
Predicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
MNI VAR DIS ENT IND COR IDM SKE STD
MNI VAR DIS ENT IND COR UNI SKE STD
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
Figure 9 Results of confusion matrix at each feature combination
refined and clearer than the thermal images on the secondrow Smoke fire and their thermal reflections are identifiedand marked in red or orange ellipses
6 Conclusion
The appropriate combination of features was investigated toaccurately classify fire smoke and their thermal reflectionsusing thermal images Gray-scale 14-bit images from a singleinfrared camera were used to extract motion and texture fea-tures by applying a clustering-based autothresholding tech-nique Bayesian classification is performed to probabilisti-cally identify multiple classes during real-time implemen-tation To find the best combination of features a multi-objective genetic algorithm optimization was implementedusing resubstitution and cross-validation errors as objectivefunctions Large-scale fire tests with different fire sources
were conducted to create a range of temperature and smokeconditions to evaluate the feature combinations
Fifteenmotion and texture features were analyzed and theprobability density functions of the features were computedby the maximum likelihood estimation The combinationof multiple features was determined to more accuratelyclassify fire smoke and thermal reflections compared witha single feature In the behavioral solution set where featurecombinations produce less than 7 resubstitution and cross-validation errors COR ENT DIS SKE STD VAR andMNIhad 800 or more occurrence while other features had400 or less occurrence The feature combination of MNIDIS COR SKE and STD produced the highest performancein the classification resulting in 668 and 670 resubstitu-tion and cross-validation errors and 9564 9761 9662and 9345 precision sensitivity F-measure and accuracyrespectively
12 Journal of Sensors
In the near future the classification of fire smoke andtheir thermal reflections will be evaluated on any classifiersand features to increase performanceThe convolution neuralnetwork of deep learning which has recently shown highperformance could be explored as a classifier also model-based image features such as discrete wavelet transform willbe further studied
Appendix
See Figure 9
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this manuscript
Acknowledgments
This work was sponsored by the Office of Naval ResearchGrant no N00014-11-1-0074 scientific office Dr ThomasMcKenna in USA Hwarang-dae Research Institute in Seouland Agency for Defense Development in Daejeon SouthKorea The authors would like to thank Mr Joseph Starr andMr JoshMcNeil for assisting in performing the fire testsTheauthors would also like to thank Rosana K Lee for helpingand supporting this research
References
[1] J-H Kim and B Y Lattimer ldquoReal-time probabilistic classifi-cation of fire and smoke using thermal imagery for intelligentfirefighting robotrdquo Fire Safety Journal vol 72 pp 40ndash49 2015
[2] J W Starr and B Y Lattimer ldquoEvaluation of navigation sensorsin fire smoke environmentsrdquo Fire Technology vol 50 no 6 pp1459ndash1481 2014
[3] J-H Kim B Keller and B Y Lattimer ldquoSensor fusion basedseek-and-find fire algorithm for intelligent firefighting robotrdquoin Proceedings of the IEEEASME International Conference onAdvanced Intelligent Mechatronics (AIM rsquo13) pp 1482ndash1486IEEE Wollongong Australia July 2013
[4] J G McNeil J Starr and B Y Lattimer ldquoAutonomous fire sup-pression using multispectral sensorsrdquo in Proceedings of theIEEEASME International Conference on Advanced IntelligentMechatronics Mechatronics for HumanWellbeing (AIM rsquo13) pp1504ndash1509 Wollongong Austrailia July 2013
[5] J-H Kim J W Starr and B Y Lattimer ldquoFirefighting robotstereo infrared vision and radar sensor fusion for imagingthrough smokerdquo Fire Technology vol 51 no 4 pp 823ndash8452015
[6] RC Luo andK L Su ldquoAutonomous fire-detection systemusingadaptive sensory fusion for intelligent security robotrdquo IEEEASME Transactions on Mechatronics vol 12 no 3 pp 274ndash2812007
[7] M A Jackson and I Robins ldquoGas sensing for fire detectionmeasurements of CO CO
2
H2
O2
and smoke density in Euro-pean standard fire testsrdquo Fire Safety Journal vol 22 no 2 pp181ndash205 1994
[8] B U Toreyin R G Cinbis Y Dedeoglu and A E Cetin ldquoFiredetection in infrared video using wavelet analysisrdquo OpticalEngineering vol 46 no 6 Article ID 067204 2007
[9] B U Toreyin Y Dedeoglu andA E Cetin ldquoWavelet based real-time smoke detection in videordquo in Proceedings of the 13th Euro-pean Signal Processing Conference pp 4ndash8 Antalya TurkeySeptember 2005
[10] T Celik H Demirel H Ozkaramanli and M Uyguroglu ldquoFiredetection using statistical color model in video sequencesrdquoJournal of Visual Communication and Image Representation vol18 no 2 pp 176ndash185 2007
[11] L Merino F Caballero J R Martınez-de-Dios I Maza and AOllero ldquoAn unmanned aircraft system for automatic forest firemonitoring and measurementrdquo Journal of Intelligent amp RoboticSystems vol 65 no 1 pp 533ndash548 2012
[12] Y Wang T W Chua R Chang and N T Pham ldquoReal-timesmoke detection using texture and color featuresrdquo in Proceed-ings of the 21st International Conference on Pattern Recognition(ICPR rsquo12) pp 1727ndash1730 Tsukuba Japan November 2012
[13] G Marbach M Loepfe and T Brupbacher ldquoAn image process-ing technique for fire detection in video imagesrdquo Fire SafetyJournal vol 41 no 4 pp 285ndash289 2006
[14] M I Chacon-Murguia and F J Perez-Vargas ldquoThermal videoanalysis for fire detection using shape regularity and intensitysaturation featuresrdquo in Pattern Recognition J F Martınez-Trinidad J A Carrasco-Ochoa C B-Y Brants and E RHancock Eds vol 6718 of Lecture Notes in Computer Sciencepp 118ndash126 Springer Berlin Germany 2011
[15] W Phillips III M Shah and N D V Lobo ldquoFlame recognitionin videordquo Pattern Recognition Letters vol 23 no 1ndash3 pp 319ndash327 2002
[16] D Han and B Lee ldquoDevelopment of early tunnel fire detectionalgorithm using the image processingrdquo in Advances in VisualComputing pp 39ndash48 Springer Berlin Germany 2006
[17] Y Chunyu F Jun W Jinjun and Z Yongming ldquoVideo firesmoke detection using motion and color featuresrdquo Fire Technol-ogy vol 46 no 3 pp 651ndash663 2010
[18] F Yuan ldquoVideo-based smoke detection with histogram se-quence of LBP and LBPV pyramidsrdquo Fire Safety Journal vol 46no 3 pp 132ndash139 2011
[19] F Lafarge X Descombes and J Zerubia ldquoTextural kernel forSVM classification in remote sensing application to forest firedetection andUrban area extractionrdquo in Proceedings of the IEEEInternational Conference on Image Processing (ICIP rsquo05) pp1096ndash1099 September 2005
[20] F Amon and A Ducharme ldquoImage frequency analysis for test-ing of fire service thermal imaging camerasrdquo Fire Technologyvol 45 no 3 pp 313ndash322 2009
[21] F Amon V Benetis J Kim and A Hamins ldquoDevelopmentof a performance evaluation facility for fire fighting thermalimagersrdquo in Defense and Security pp 244ndash252 2004
[22] F D Maxwell ldquoA portable IR system for observing fire thrusmokerdquo Fire Technology vol 7 no 4 pp 321ndash331 1971
[23] A Barducci D Guzzi P Marcoionni and I Pippi ldquoInfrareddetection of active fires and burnt areas theory and observa-tionsrdquo Infrared Physics amp Technology vol 43 no 3ndash5 pp 119ndash125 2002
[24] C Wang and S Qin ldquoAdaptive detection method of infraredsmall target based on target-background separation via robustprincipal component analysisrdquo Infrared Physics amp Technologyvol 69 pp 123ndash135 2015
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
8 Journal of Sensors
Table 5 Results of error case and performance at each feature combination in the behavioral solution set (Resu means resubstitution andCros refers to cross-validation)
Index Combination of features Error () Case Performance ()Resu Cros TP FP FN Precision Sensitivity 119865-measure Accuracy
1 MNI VAR ENT COR SKE 690 689 10049 402 324 9615 9688 9651 93262 MNI DIS COR SKE STD 668 670 10069 459 247 9564 9761 9662 93453 MNI ENT COR SKE STD 676 675 10061 447 267 9575 9741 9657 9337
4 MNI VAR DIS ENT CON CORSKE STD 698 690 9980 454 341 9565 9670 9617 9262
5 MNI VAR DIS ENT COR UNISKE STD 689 690 10037 453 285 9568 9724 9645 9315
6 MNI VAR DIS ENT COR IDMSKE STD 677 679 10047 461 267 9561 9741 9650 9324
7 MNI VAR DIS ENT IDM SKEKUR STD 696 694 10033 505 237 9521 9769 9643 9311
8 MNI VAR DIS ENT IND CORUNI SKE STD 695 699 10029 451 295 9570 9714 9641 9308
9 MNI VAR DIS ENT IND CORIDM SKE STD 688 690 10035 457 283 9564 9726 9644 9313
10 MNI VAR DIS ENT COR IDMSKE KUR STD 689 691 10036 478 261 9545 9747 9645 9314
PrecisionSensitivity
F-measureAccuracy
90919293949596979899
Perfo
rman
ce (
)
2 3 4 5 6 7 8 9 101Feature combination index
Figure 6 Results of performance at each feature combination in thebehavioral solution set
to distinguish The best solution was determined to be indexnumber 2 combination of MNI DIS COR SKE and STDwhich has the lowest of resubstitution and cross-validationerrors 668 and 670 respectively This combinationincludes all of the top features based on the probability occur-rence except ENT and VAR The four performance results ateach feature combination in the behavioral solution set areshown in Figure 6 where the highest results aremarked in redcircles and the lowest in green-dot circles Sensitivity appearshigher than precision at each feature combination becauseFPs are larger than FNs in the confusion matrix Particularlyindex number 7 has the biggest difference between FP and FNresulting in the highest sensitivity and lowest precision Thesummation of FP and FN in index number 4 is the highestin the behavioral solution set resulting in the lowest accuracywhile index number 2 has the lowest summation of FP andFN providing the highest accuracy
This study investigated a wide range of features fromlong-wavelength infrared camera images analyzed normaldistributions of fifteen features with respect to the classes ofsmoke fire and their thermal reflections and discovered thehighest performing feature combination by examining singlefeatures and multiple feature combinations As a result theproposed feature combination of MNI DIS COR SKE andSTD increases the performance compared with the previousstudy [1] which used MNI VAR ENT and IDM As shownin Figure 7 the errors are reduced by 286 and 268 resub-stitution and cross-validation errors and performances areincreased by 290 158 020 and 285 accuracy F-measure sensitivity and precision respectively
Figure 8 shows original visual and thermal images withthe robot at three different locations start point hallwayentrance and room entrance described in the experimentalfacility Each row relates to a series of images from the robotat three locations The first row contains visible images ofthe robot view As seen in the visible image at start pointfurther information regarding the hallway is limited due toshadowing of the light The image at hallway entrance showsa smoke layer in the upper portion of the hallway due to afire inside the roomThe image at the room entrance displaysa wood crib fire with sparks Because of soot and relativedifference in brightness the background is shown darker andthus limiting information on the background around the fire
Thermal infrared images are displayed in the second rowto show information that RGB camera cannot provide infire environments Unlike visual image at start point thatis obscured due to shadowing the presence of smoke andits thermal reflections on the ventilation hood can be obvi-ously perceived The red boxes on thermal images indicateobjects extracted through the adaptive object extraction withoptical flows and identification numbers In spite of dense
Journal of Sensors 9
9279
9741
9504
9055
954
938
9564
9761
9662
9345
668
670
Precision
Sensitivity
F-measure
Accuracy
Resubstitution error
Cross-validation error
The proposedThe previous
2000 4000 6000 8000 10000000()
Figure 7 Result comparison between the previous and the proposed studies
① Start point ② Hallway entrance ③ Room entrance
IR im
age
class
ifica
tion
resu
lts
Visib
le im
age
IR im
age
imag
e ana
lysis
Figure 8 Original visible and IR images with image analysis and classification results at different locations in the test setup of actual fires
smoke-filled and low visibility environments thermal imagescan generate the images of smoke and fire as well as back-ground information that is otherwise not visible throughvisual imaging
On the third row class labels and posterior probabilitiesof each candidate are displayed at the center of candidateROI as a result of Bayesian classification Using enhancedimage processing techniques the thermal images can bemore
10 Journal of Sensors
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
2
0
19
4820
64
Feature combination index number 2
Feature combination index number 3 Feature combination index number 4
Feature combination index number 5
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1462
0
0
0
1
1462
0
0
0
1
1457
0
0
0
1
1457
0
0
1365
193
197
0
0
1343
181
196
0
0
1325
205
181
0
0
1270
207
181
0
0
53
167
1929
0
0
145
166
1943
0
0
101
165
1942
7
0
49
4844
62
2
0
30
4817
63
0
0
53
165
1929
7
0
27
4830
61
MNI DIS COR SKE STD
MNI ENT COR SKE STD MNI VAR DIS ENT CON COR SKE STD
MNI VAR DIS ENT COR UNI SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bjPredicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Feature combination index number 1
140
159
0
0
0
0
0
00
0
0
0
0
0
1981
488
1280
193
144
1
1462
4838
62
25
2
MNI VAR ENT COR SKE
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Feature combination index number 6
0
0
0
0
488
0
0
0
1
14620
0
1339
206
185
0
0
93
161
1935
2
0
13
4823
67
MNI VAR DIS ENT COR IDM SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
F-re
flect
ion
Oth
er o
bj
Smok
e
Fire
S-re
flect
ion
Predicted class
Figure 9 Continued
Journal of Sensors 11
Feature combination index number 7 Feature combination index number 8
Feature combination index number 9 Feature combination index number 10
89
133
0
0
0
0
0
02
0
0
0
0
0
1888
488
0
0
0
0
488
1341
200
235
1
1459
4857
62
15
5
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1458
0
0
0
1
1462
0
0
0
1
1462
0
0
1324
204
181
0
0
1334
208
181
0
0
1334
205
204
0
0
99
174
1943
0
0
91
150
1921
2
0
18
4810
65
6
0
20
4835
62
0
0
93
172
1941
2
0
22
4812
63
MNI VAR DIS ENT COR IDM SKE KUR STD
MNI VAR DIS ENT IDM SKE KUR STD
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bj
Predicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
MNI VAR DIS ENT IND COR IDM SKE STD
MNI VAR DIS ENT IND COR UNI SKE STD
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
Figure 9 Results of confusion matrix at each feature combination
refined and clearer than the thermal images on the secondrow Smoke fire and their thermal reflections are identifiedand marked in red or orange ellipses
6 Conclusion
The appropriate combination of features was investigated toaccurately classify fire smoke and their thermal reflectionsusing thermal images Gray-scale 14-bit images from a singleinfrared camera were used to extract motion and texture fea-tures by applying a clustering-based autothresholding tech-nique Bayesian classification is performed to probabilisti-cally identify multiple classes during real-time implemen-tation To find the best combination of features a multi-objective genetic algorithm optimization was implementedusing resubstitution and cross-validation errors as objectivefunctions Large-scale fire tests with different fire sources
were conducted to create a range of temperature and smokeconditions to evaluate the feature combinations
Fifteenmotion and texture features were analyzed and theprobability density functions of the features were computedby the maximum likelihood estimation The combinationof multiple features was determined to more accuratelyclassify fire smoke and thermal reflections compared witha single feature In the behavioral solution set where featurecombinations produce less than 7 resubstitution and cross-validation errors COR ENT DIS SKE STD VAR andMNIhad 800 or more occurrence while other features had400 or less occurrence The feature combination of MNIDIS COR SKE and STD produced the highest performancein the classification resulting in 668 and 670 resubstitu-tion and cross-validation errors and 9564 9761 9662and 9345 precision sensitivity F-measure and accuracyrespectively
12 Journal of Sensors
In the near future the classification of fire smoke andtheir thermal reflections will be evaluated on any classifiersand features to increase performanceThe convolution neuralnetwork of deep learning which has recently shown highperformance could be explored as a classifier also model-based image features such as discrete wavelet transform willbe further studied
Appendix
See Figure 9
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this manuscript
Acknowledgments
This work was sponsored by the Office of Naval ResearchGrant no N00014-11-1-0074 scientific office Dr ThomasMcKenna in USA Hwarang-dae Research Institute in Seouland Agency for Defense Development in Daejeon SouthKorea The authors would like to thank Mr Joseph Starr andMr JoshMcNeil for assisting in performing the fire testsTheauthors would also like to thank Rosana K Lee for helpingand supporting this research
References
[1] J-H Kim and B Y Lattimer ldquoReal-time probabilistic classifi-cation of fire and smoke using thermal imagery for intelligentfirefighting robotrdquo Fire Safety Journal vol 72 pp 40ndash49 2015
[2] J W Starr and B Y Lattimer ldquoEvaluation of navigation sensorsin fire smoke environmentsrdquo Fire Technology vol 50 no 6 pp1459ndash1481 2014
[3] J-H Kim B Keller and B Y Lattimer ldquoSensor fusion basedseek-and-find fire algorithm for intelligent firefighting robotrdquoin Proceedings of the IEEEASME International Conference onAdvanced Intelligent Mechatronics (AIM rsquo13) pp 1482ndash1486IEEE Wollongong Australia July 2013
[4] J G McNeil J Starr and B Y Lattimer ldquoAutonomous fire sup-pression using multispectral sensorsrdquo in Proceedings of theIEEEASME International Conference on Advanced IntelligentMechatronics Mechatronics for HumanWellbeing (AIM rsquo13) pp1504ndash1509 Wollongong Austrailia July 2013
[5] J-H Kim J W Starr and B Y Lattimer ldquoFirefighting robotstereo infrared vision and radar sensor fusion for imagingthrough smokerdquo Fire Technology vol 51 no 4 pp 823ndash8452015
[6] RC Luo andK L Su ldquoAutonomous fire-detection systemusingadaptive sensory fusion for intelligent security robotrdquo IEEEASME Transactions on Mechatronics vol 12 no 3 pp 274ndash2812007
[7] M A Jackson and I Robins ldquoGas sensing for fire detectionmeasurements of CO CO
2
H2
O2
and smoke density in Euro-pean standard fire testsrdquo Fire Safety Journal vol 22 no 2 pp181ndash205 1994
[8] B U Toreyin R G Cinbis Y Dedeoglu and A E Cetin ldquoFiredetection in infrared video using wavelet analysisrdquo OpticalEngineering vol 46 no 6 Article ID 067204 2007
[9] B U Toreyin Y Dedeoglu andA E Cetin ldquoWavelet based real-time smoke detection in videordquo in Proceedings of the 13th Euro-pean Signal Processing Conference pp 4ndash8 Antalya TurkeySeptember 2005
[10] T Celik H Demirel H Ozkaramanli and M Uyguroglu ldquoFiredetection using statistical color model in video sequencesrdquoJournal of Visual Communication and Image Representation vol18 no 2 pp 176ndash185 2007
[11] L Merino F Caballero J R Martınez-de-Dios I Maza and AOllero ldquoAn unmanned aircraft system for automatic forest firemonitoring and measurementrdquo Journal of Intelligent amp RoboticSystems vol 65 no 1 pp 533ndash548 2012
[12] Y Wang T W Chua R Chang and N T Pham ldquoReal-timesmoke detection using texture and color featuresrdquo in Proceed-ings of the 21st International Conference on Pattern Recognition(ICPR rsquo12) pp 1727ndash1730 Tsukuba Japan November 2012
[13] G Marbach M Loepfe and T Brupbacher ldquoAn image process-ing technique for fire detection in video imagesrdquo Fire SafetyJournal vol 41 no 4 pp 285ndash289 2006
[14] M I Chacon-Murguia and F J Perez-Vargas ldquoThermal videoanalysis for fire detection using shape regularity and intensitysaturation featuresrdquo in Pattern Recognition J F Martınez-Trinidad J A Carrasco-Ochoa C B-Y Brants and E RHancock Eds vol 6718 of Lecture Notes in Computer Sciencepp 118ndash126 Springer Berlin Germany 2011
[15] W Phillips III M Shah and N D V Lobo ldquoFlame recognitionin videordquo Pattern Recognition Letters vol 23 no 1ndash3 pp 319ndash327 2002
[16] D Han and B Lee ldquoDevelopment of early tunnel fire detectionalgorithm using the image processingrdquo in Advances in VisualComputing pp 39ndash48 Springer Berlin Germany 2006
[17] Y Chunyu F Jun W Jinjun and Z Yongming ldquoVideo firesmoke detection using motion and color featuresrdquo Fire Technol-ogy vol 46 no 3 pp 651ndash663 2010
[18] F Yuan ldquoVideo-based smoke detection with histogram se-quence of LBP and LBPV pyramidsrdquo Fire Safety Journal vol 46no 3 pp 132ndash139 2011
[19] F Lafarge X Descombes and J Zerubia ldquoTextural kernel forSVM classification in remote sensing application to forest firedetection andUrban area extractionrdquo in Proceedings of the IEEEInternational Conference on Image Processing (ICIP rsquo05) pp1096ndash1099 September 2005
[20] F Amon and A Ducharme ldquoImage frequency analysis for test-ing of fire service thermal imaging camerasrdquo Fire Technologyvol 45 no 3 pp 313ndash322 2009
[21] F Amon V Benetis J Kim and A Hamins ldquoDevelopmentof a performance evaluation facility for fire fighting thermalimagersrdquo in Defense and Security pp 244ndash252 2004
[22] F D Maxwell ldquoA portable IR system for observing fire thrusmokerdquo Fire Technology vol 7 no 4 pp 321ndash331 1971
[23] A Barducci D Guzzi P Marcoionni and I Pippi ldquoInfrareddetection of active fires and burnt areas theory and observa-tionsrdquo Infrared Physics amp Technology vol 43 no 3ndash5 pp 119ndash125 2002
[24] C Wang and S Qin ldquoAdaptive detection method of infraredsmall target based on target-background separation via robustprincipal component analysisrdquo Infrared Physics amp Technologyvol 69 pp 123ndash135 2015
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
Journal of Sensors 9
9279
9741
9504
9055
954
938
9564
9761
9662
9345
668
670
Precision
Sensitivity
F-measure
Accuracy
Resubstitution error
Cross-validation error
The proposedThe previous
2000 4000 6000 8000 10000000()
Figure 7 Result comparison between the previous and the proposed studies
① Start point ② Hallway entrance ③ Room entrance
IR im
age
class
ifica
tion
resu
lts
Visib
le im
age
IR im
age
imag
e ana
lysis
Figure 8 Original visible and IR images with image analysis and classification results at different locations in the test setup of actual fires
smoke-filled and low visibility environments thermal imagescan generate the images of smoke and fire as well as back-ground information that is otherwise not visible throughvisual imaging
On the third row class labels and posterior probabilitiesof each candidate are displayed at the center of candidateROI as a result of Bayesian classification Using enhancedimage processing techniques the thermal images can bemore
10 Journal of Sensors
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
2
0
19
4820
64
Feature combination index number 2
Feature combination index number 3 Feature combination index number 4
Feature combination index number 5
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1462
0
0
0
1
1462
0
0
0
1
1457
0
0
0
1
1457
0
0
1365
193
197
0
0
1343
181
196
0
0
1325
205
181
0
0
1270
207
181
0
0
53
167
1929
0
0
145
166
1943
0
0
101
165
1942
7
0
49
4844
62
2
0
30
4817
63
0
0
53
165
1929
7
0
27
4830
61
MNI DIS COR SKE STD
MNI ENT COR SKE STD MNI VAR DIS ENT CON COR SKE STD
MNI VAR DIS ENT COR UNI SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bjPredicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Feature combination index number 1
140
159
0
0
0
0
0
00
0
0
0
0
0
1981
488
1280
193
144
1
1462
4838
62
25
2
MNI VAR ENT COR SKE
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Feature combination index number 6
0
0
0
0
488
0
0
0
1
14620
0
1339
206
185
0
0
93
161
1935
2
0
13
4823
67
MNI VAR DIS ENT COR IDM SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
F-re
flect
ion
Oth
er o
bj
Smok
e
Fire
S-re
flect
ion
Predicted class
Figure 9 Continued
Journal of Sensors 11
Feature combination index number 7 Feature combination index number 8
Feature combination index number 9 Feature combination index number 10
89
133
0
0
0
0
0
02
0
0
0
0
0
1888
488
0
0
0
0
488
1341
200
235
1
1459
4857
62
15
5
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1458
0
0
0
1
1462
0
0
0
1
1462
0
0
1324
204
181
0
0
1334
208
181
0
0
1334
205
204
0
0
99
174
1943
0
0
91
150
1921
2
0
18
4810
65
6
0
20
4835
62
0
0
93
172
1941
2
0
22
4812
63
MNI VAR DIS ENT COR IDM SKE KUR STD
MNI VAR DIS ENT IDM SKE KUR STD
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bj
Predicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
MNI VAR DIS ENT IND COR IDM SKE STD
MNI VAR DIS ENT IND COR UNI SKE STD
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
Figure 9 Results of confusion matrix at each feature combination
refined and clearer than the thermal images on the secondrow Smoke fire and their thermal reflections are identifiedand marked in red or orange ellipses
6 Conclusion
The appropriate combination of features was investigated toaccurately classify fire smoke and their thermal reflectionsusing thermal images Gray-scale 14-bit images from a singleinfrared camera were used to extract motion and texture fea-tures by applying a clustering-based autothresholding tech-nique Bayesian classification is performed to probabilisti-cally identify multiple classes during real-time implemen-tation To find the best combination of features a multi-objective genetic algorithm optimization was implementedusing resubstitution and cross-validation errors as objectivefunctions Large-scale fire tests with different fire sources
were conducted to create a range of temperature and smokeconditions to evaluate the feature combinations
Fifteenmotion and texture features were analyzed and theprobability density functions of the features were computedby the maximum likelihood estimation The combinationof multiple features was determined to more accuratelyclassify fire smoke and thermal reflections compared witha single feature In the behavioral solution set where featurecombinations produce less than 7 resubstitution and cross-validation errors COR ENT DIS SKE STD VAR andMNIhad 800 or more occurrence while other features had400 or less occurrence The feature combination of MNIDIS COR SKE and STD produced the highest performancein the classification resulting in 668 and 670 resubstitu-tion and cross-validation errors and 9564 9761 9662and 9345 precision sensitivity F-measure and accuracyrespectively
12 Journal of Sensors
In the near future the classification of fire smoke andtheir thermal reflections will be evaluated on any classifiersand features to increase performanceThe convolution neuralnetwork of deep learning which has recently shown highperformance could be explored as a classifier also model-based image features such as discrete wavelet transform willbe further studied
Appendix
See Figure 9
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this manuscript
Acknowledgments
This work was sponsored by the Office of Naval ResearchGrant no N00014-11-1-0074 scientific office Dr ThomasMcKenna in USA Hwarang-dae Research Institute in Seouland Agency for Defense Development in Daejeon SouthKorea The authors would like to thank Mr Joseph Starr andMr JoshMcNeil for assisting in performing the fire testsTheauthors would also like to thank Rosana K Lee for helpingand supporting this research
References
[1] J-H Kim and B Y Lattimer ldquoReal-time probabilistic classifi-cation of fire and smoke using thermal imagery for intelligentfirefighting robotrdquo Fire Safety Journal vol 72 pp 40ndash49 2015
[2] J W Starr and B Y Lattimer ldquoEvaluation of navigation sensorsin fire smoke environmentsrdquo Fire Technology vol 50 no 6 pp1459ndash1481 2014
[3] J-H Kim B Keller and B Y Lattimer ldquoSensor fusion basedseek-and-find fire algorithm for intelligent firefighting robotrdquoin Proceedings of the IEEEASME International Conference onAdvanced Intelligent Mechatronics (AIM rsquo13) pp 1482ndash1486IEEE Wollongong Australia July 2013
[4] J G McNeil J Starr and B Y Lattimer ldquoAutonomous fire sup-pression using multispectral sensorsrdquo in Proceedings of theIEEEASME International Conference on Advanced IntelligentMechatronics Mechatronics for HumanWellbeing (AIM rsquo13) pp1504ndash1509 Wollongong Austrailia July 2013
[5] J-H Kim J W Starr and B Y Lattimer ldquoFirefighting robotstereo infrared vision and radar sensor fusion for imagingthrough smokerdquo Fire Technology vol 51 no 4 pp 823ndash8452015
[6] RC Luo andK L Su ldquoAutonomous fire-detection systemusingadaptive sensory fusion for intelligent security robotrdquo IEEEASME Transactions on Mechatronics vol 12 no 3 pp 274ndash2812007
[7] M A Jackson and I Robins ldquoGas sensing for fire detectionmeasurements of CO CO
2
H2
O2
and smoke density in Euro-pean standard fire testsrdquo Fire Safety Journal vol 22 no 2 pp181ndash205 1994
[8] B U Toreyin R G Cinbis Y Dedeoglu and A E Cetin ldquoFiredetection in infrared video using wavelet analysisrdquo OpticalEngineering vol 46 no 6 Article ID 067204 2007
[9] B U Toreyin Y Dedeoglu andA E Cetin ldquoWavelet based real-time smoke detection in videordquo in Proceedings of the 13th Euro-pean Signal Processing Conference pp 4ndash8 Antalya TurkeySeptember 2005
[10] T Celik H Demirel H Ozkaramanli and M Uyguroglu ldquoFiredetection using statistical color model in video sequencesrdquoJournal of Visual Communication and Image Representation vol18 no 2 pp 176ndash185 2007
[11] L Merino F Caballero J R Martınez-de-Dios I Maza and AOllero ldquoAn unmanned aircraft system for automatic forest firemonitoring and measurementrdquo Journal of Intelligent amp RoboticSystems vol 65 no 1 pp 533ndash548 2012
[12] Y Wang T W Chua R Chang and N T Pham ldquoReal-timesmoke detection using texture and color featuresrdquo in Proceed-ings of the 21st International Conference on Pattern Recognition(ICPR rsquo12) pp 1727ndash1730 Tsukuba Japan November 2012
[13] G Marbach M Loepfe and T Brupbacher ldquoAn image process-ing technique for fire detection in video imagesrdquo Fire SafetyJournal vol 41 no 4 pp 285ndash289 2006
[14] M I Chacon-Murguia and F J Perez-Vargas ldquoThermal videoanalysis for fire detection using shape regularity and intensitysaturation featuresrdquo in Pattern Recognition J F Martınez-Trinidad J A Carrasco-Ochoa C B-Y Brants and E RHancock Eds vol 6718 of Lecture Notes in Computer Sciencepp 118ndash126 Springer Berlin Germany 2011
[15] W Phillips III M Shah and N D V Lobo ldquoFlame recognitionin videordquo Pattern Recognition Letters vol 23 no 1ndash3 pp 319ndash327 2002
[16] D Han and B Lee ldquoDevelopment of early tunnel fire detectionalgorithm using the image processingrdquo in Advances in VisualComputing pp 39ndash48 Springer Berlin Germany 2006
[17] Y Chunyu F Jun W Jinjun and Z Yongming ldquoVideo firesmoke detection using motion and color featuresrdquo Fire Technol-ogy vol 46 no 3 pp 651ndash663 2010
[18] F Yuan ldquoVideo-based smoke detection with histogram se-quence of LBP and LBPV pyramidsrdquo Fire Safety Journal vol 46no 3 pp 132ndash139 2011
[19] F Lafarge X Descombes and J Zerubia ldquoTextural kernel forSVM classification in remote sensing application to forest firedetection andUrban area extractionrdquo in Proceedings of the IEEEInternational Conference on Image Processing (ICIP rsquo05) pp1096ndash1099 September 2005
[20] F Amon and A Ducharme ldquoImage frequency analysis for test-ing of fire service thermal imaging camerasrdquo Fire Technologyvol 45 no 3 pp 313ndash322 2009
[21] F Amon V Benetis J Kim and A Hamins ldquoDevelopmentof a performance evaluation facility for fire fighting thermalimagersrdquo in Defense and Security pp 244ndash252 2004
[22] F D Maxwell ldquoA portable IR system for observing fire thrusmokerdquo Fire Technology vol 7 no 4 pp 321ndash331 1971
[23] A Barducci D Guzzi P Marcoionni and I Pippi ldquoInfrareddetection of active fires and burnt areas theory and observa-tionsrdquo Infrared Physics amp Technology vol 43 no 3ndash5 pp 119ndash125 2002
[24] C Wang and S Qin ldquoAdaptive detection method of infraredsmall target based on target-background separation via robustprincipal component analysisrdquo Infrared Physics amp Technologyvol 69 pp 123ndash135 2015
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
10 Journal of Sensors
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
2
0
19
4820
64
Feature combination index number 2
Feature combination index number 3 Feature combination index number 4
Feature combination index number 5
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1462
0
0
0
1
1462
0
0
0
1
1457
0
0
0
1
1457
0
0
1365
193
197
0
0
1343
181
196
0
0
1325
205
181
0
0
1270
207
181
0
0
53
167
1929
0
0
145
166
1943
0
0
101
165
1942
7
0
49
4844
62
2
0
30
4817
63
0
0
53
165
1929
7
0
27
4830
61
MNI DIS COR SKE STD
MNI ENT COR SKE STD MNI VAR DIS ENT CON COR SKE STD
MNI VAR DIS ENT COR UNI SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bjPredicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Feature combination index number 1
140
159
0
0
0
0
0
00
0
0
0
0
0
1981
488
1280
193
144
1
1462
4838
62
25
2
MNI VAR ENT COR SKE
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Feature combination index number 6
0
0
0
0
488
0
0
0
1
14620
0
1339
206
185
0
0
93
161
1935
2
0
13
4823
67
MNI VAR DIS ENT COR IDM SKE STD
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
F-re
flect
ion
Oth
er o
bj
Smok
e
Fire
S-re
flect
ion
Predicted class
Figure 9 Continued
Journal of Sensors 11
Feature combination index number 7 Feature combination index number 8
Feature combination index number 9 Feature combination index number 10
89
133
0
0
0
0
0
02
0
0
0
0
0
1888
488
0
0
0
0
488
1341
200
235
1
1459
4857
62
15
5
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1458
0
0
0
1
1462
0
0
0
1
1462
0
0
1324
204
181
0
0
1334
208
181
0
0
1334
205
204
0
0
99
174
1943
0
0
91
150
1921
2
0
18
4810
65
6
0
20
4835
62
0
0
93
172
1941
2
0
22
4812
63
MNI VAR DIS ENT COR IDM SKE KUR STD
MNI VAR DIS ENT IDM SKE KUR STD
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bj
Predicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
MNI VAR DIS ENT IND COR IDM SKE STD
MNI VAR DIS ENT IND COR UNI SKE STD
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
Figure 9 Results of confusion matrix at each feature combination
refined and clearer than the thermal images on the secondrow Smoke fire and their thermal reflections are identifiedand marked in red or orange ellipses
6 Conclusion
The appropriate combination of features was investigated toaccurately classify fire smoke and their thermal reflectionsusing thermal images Gray-scale 14-bit images from a singleinfrared camera were used to extract motion and texture fea-tures by applying a clustering-based autothresholding tech-nique Bayesian classification is performed to probabilisti-cally identify multiple classes during real-time implemen-tation To find the best combination of features a multi-objective genetic algorithm optimization was implementedusing resubstitution and cross-validation errors as objectivefunctions Large-scale fire tests with different fire sources
were conducted to create a range of temperature and smokeconditions to evaluate the feature combinations
Fifteenmotion and texture features were analyzed and theprobability density functions of the features were computedby the maximum likelihood estimation The combinationof multiple features was determined to more accuratelyclassify fire smoke and thermal reflections compared witha single feature In the behavioral solution set where featurecombinations produce less than 7 resubstitution and cross-validation errors COR ENT DIS SKE STD VAR andMNIhad 800 or more occurrence while other features had400 or less occurrence The feature combination of MNIDIS COR SKE and STD produced the highest performancein the classification resulting in 668 and 670 resubstitu-tion and cross-validation errors and 9564 9761 9662and 9345 precision sensitivity F-measure and accuracyrespectively
12 Journal of Sensors
In the near future the classification of fire smoke andtheir thermal reflections will be evaluated on any classifiersand features to increase performanceThe convolution neuralnetwork of deep learning which has recently shown highperformance could be explored as a classifier also model-based image features such as discrete wavelet transform willbe further studied
Appendix
See Figure 9
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this manuscript
Acknowledgments
This work was sponsored by the Office of Naval ResearchGrant no N00014-11-1-0074 scientific office Dr ThomasMcKenna in USA Hwarang-dae Research Institute in Seouland Agency for Defense Development in Daejeon SouthKorea The authors would like to thank Mr Joseph Starr andMr JoshMcNeil for assisting in performing the fire testsTheauthors would also like to thank Rosana K Lee for helpingand supporting this research
References
[1] J-H Kim and B Y Lattimer ldquoReal-time probabilistic classifi-cation of fire and smoke using thermal imagery for intelligentfirefighting robotrdquo Fire Safety Journal vol 72 pp 40ndash49 2015
[2] J W Starr and B Y Lattimer ldquoEvaluation of navigation sensorsin fire smoke environmentsrdquo Fire Technology vol 50 no 6 pp1459ndash1481 2014
[3] J-H Kim B Keller and B Y Lattimer ldquoSensor fusion basedseek-and-find fire algorithm for intelligent firefighting robotrdquoin Proceedings of the IEEEASME International Conference onAdvanced Intelligent Mechatronics (AIM rsquo13) pp 1482ndash1486IEEE Wollongong Australia July 2013
[4] J G McNeil J Starr and B Y Lattimer ldquoAutonomous fire sup-pression using multispectral sensorsrdquo in Proceedings of theIEEEASME International Conference on Advanced IntelligentMechatronics Mechatronics for HumanWellbeing (AIM rsquo13) pp1504ndash1509 Wollongong Austrailia July 2013
[5] J-H Kim J W Starr and B Y Lattimer ldquoFirefighting robotstereo infrared vision and radar sensor fusion for imagingthrough smokerdquo Fire Technology vol 51 no 4 pp 823ndash8452015
[6] RC Luo andK L Su ldquoAutonomous fire-detection systemusingadaptive sensory fusion for intelligent security robotrdquo IEEEASME Transactions on Mechatronics vol 12 no 3 pp 274ndash2812007
[7] M A Jackson and I Robins ldquoGas sensing for fire detectionmeasurements of CO CO
2
H2
O2
and smoke density in Euro-pean standard fire testsrdquo Fire Safety Journal vol 22 no 2 pp181ndash205 1994
[8] B U Toreyin R G Cinbis Y Dedeoglu and A E Cetin ldquoFiredetection in infrared video using wavelet analysisrdquo OpticalEngineering vol 46 no 6 Article ID 067204 2007
[9] B U Toreyin Y Dedeoglu andA E Cetin ldquoWavelet based real-time smoke detection in videordquo in Proceedings of the 13th Euro-pean Signal Processing Conference pp 4ndash8 Antalya TurkeySeptember 2005
[10] T Celik H Demirel H Ozkaramanli and M Uyguroglu ldquoFiredetection using statistical color model in video sequencesrdquoJournal of Visual Communication and Image Representation vol18 no 2 pp 176ndash185 2007
[11] L Merino F Caballero J R Martınez-de-Dios I Maza and AOllero ldquoAn unmanned aircraft system for automatic forest firemonitoring and measurementrdquo Journal of Intelligent amp RoboticSystems vol 65 no 1 pp 533ndash548 2012
[12] Y Wang T W Chua R Chang and N T Pham ldquoReal-timesmoke detection using texture and color featuresrdquo in Proceed-ings of the 21st International Conference on Pattern Recognition(ICPR rsquo12) pp 1727ndash1730 Tsukuba Japan November 2012
[13] G Marbach M Loepfe and T Brupbacher ldquoAn image process-ing technique for fire detection in video imagesrdquo Fire SafetyJournal vol 41 no 4 pp 285ndash289 2006
[14] M I Chacon-Murguia and F J Perez-Vargas ldquoThermal videoanalysis for fire detection using shape regularity and intensitysaturation featuresrdquo in Pattern Recognition J F Martınez-Trinidad J A Carrasco-Ochoa C B-Y Brants and E RHancock Eds vol 6718 of Lecture Notes in Computer Sciencepp 118ndash126 Springer Berlin Germany 2011
[15] W Phillips III M Shah and N D V Lobo ldquoFlame recognitionin videordquo Pattern Recognition Letters vol 23 no 1ndash3 pp 319ndash327 2002
[16] D Han and B Lee ldquoDevelopment of early tunnel fire detectionalgorithm using the image processingrdquo in Advances in VisualComputing pp 39ndash48 Springer Berlin Germany 2006
[17] Y Chunyu F Jun W Jinjun and Z Yongming ldquoVideo firesmoke detection using motion and color featuresrdquo Fire Technol-ogy vol 46 no 3 pp 651ndash663 2010
[18] F Yuan ldquoVideo-based smoke detection with histogram se-quence of LBP and LBPV pyramidsrdquo Fire Safety Journal vol 46no 3 pp 132ndash139 2011
[19] F Lafarge X Descombes and J Zerubia ldquoTextural kernel forSVM classification in remote sensing application to forest firedetection andUrban area extractionrdquo in Proceedings of the IEEEInternational Conference on Image Processing (ICIP rsquo05) pp1096ndash1099 September 2005
[20] F Amon and A Ducharme ldquoImage frequency analysis for test-ing of fire service thermal imaging camerasrdquo Fire Technologyvol 45 no 3 pp 313ndash322 2009
[21] F Amon V Benetis J Kim and A Hamins ldquoDevelopmentof a performance evaluation facility for fire fighting thermalimagersrdquo in Defense and Security pp 244ndash252 2004
[22] F D Maxwell ldquoA portable IR system for observing fire thrusmokerdquo Fire Technology vol 7 no 4 pp 321ndash331 1971
[23] A Barducci D Guzzi P Marcoionni and I Pippi ldquoInfrareddetection of active fires and burnt areas theory and observa-tionsrdquo Infrared Physics amp Technology vol 43 no 3ndash5 pp 119ndash125 2002
[24] C Wang and S Qin ldquoAdaptive detection method of infraredsmall target based on target-background separation via robustprincipal component analysisrdquo Infrared Physics amp Technologyvol 69 pp 123ndash135 2015
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
Journal of Sensors 11
Feature combination index number 7 Feature combination index number 8
Feature combination index number 9 Feature combination index number 10
89
133
0
0
0
0
0
02
0
0
0
0
0
1888
488
0
0
0
0
488
1341
200
235
1
1459
4857
62
15
5
0
0
0
0
488
0
0
0
0
488
0
0
0
1
1458
0
0
0
1
1462
0
0
0
1
1462
0
0
1324
204
181
0
0
1334
208
181
0
0
1334
205
204
0
0
99
174
1943
0
0
91
150
1921
2
0
18
4810
65
6
0
20
4835
62
0
0
93
172
1941
2
0
22
4812
63
MNI VAR DIS ENT COR IDM SKE KUR STD
MNI VAR DIS ENT IDM SKE KUR STD
Smok
e
S-re
flect
ion
Fire
F-re
flect
ion
Oth
er o
bj
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflectionAc
tual
clas
s
Smok
e
F-re
flect
ion
Fire
S-re
flect
ion
Oth
er o
bj
Predicted class
Smok
e
Fire
S-re
flect
ion
Oth
er o
bj
F-re
flect
ion
Predicted class
Other obj
F-reflection
Fire
Smoke
S-reflection
Actu
al cl
ass
MNI VAR DIS ENT IND COR IDM SKE STD
MNI VAR DIS ENT IND COR UNI SKE STD
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
050010001500200025003000350040004500
Figure 9 Results of confusion matrix at each feature combination
refined and clearer than the thermal images on the secondrow Smoke fire and their thermal reflections are identifiedand marked in red or orange ellipses
6 Conclusion
The appropriate combination of features was investigated toaccurately classify fire smoke and their thermal reflectionsusing thermal images Gray-scale 14-bit images from a singleinfrared camera were used to extract motion and texture fea-tures by applying a clustering-based autothresholding tech-nique Bayesian classification is performed to probabilisti-cally identify multiple classes during real-time implemen-tation To find the best combination of features a multi-objective genetic algorithm optimization was implementedusing resubstitution and cross-validation errors as objectivefunctions Large-scale fire tests with different fire sources
were conducted to create a range of temperature and smokeconditions to evaluate the feature combinations
Fifteenmotion and texture features were analyzed and theprobability density functions of the features were computedby the maximum likelihood estimation The combinationof multiple features was determined to more accuratelyclassify fire smoke and thermal reflections compared witha single feature In the behavioral solution set where featurecombinations produce less than 7 resubstitution and cross-validation errors COR ENT DIS SKE STD VAR andMNIhad 800 or more occurrence while other features had400 or less occurrence The feature combination of MNIDIS COR SKE and STD produced the highest performancein the classification resulting in 668 and 670 resubstitu-tion and cross-validation errors and 9564 9761 9662and 9345 precision sensitivity F-measure and accuracyrespectively
12 Journal of Sensors
In the near future the classification of fire smoke andtheir thermal reflections will be evaluated on any classifiersand features to increase performanceThe convolution neuralnetwork of deep learning which has recently shown highperformance could be explored as a classifier also model-based image features such as discrete wavelet transform willbe further studied
Appendix
See Figure 9
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this manuscript
Acknowledgments
This work was sponsored by the Office of Naval ResearchGrant no N00014-11-1-0074 scientific office Dr ThomasMcKenna in USA Hwarang-dae Research Institute in Seouland Agency for Defense Development in Daejeon SouthKorea The authors would like to thank Mr Joseph Starr andMr JoshMcNeil for assisting in performing the fire testsTheauthors would also like to thank Rosana K Lee for helpingand supporting this research
References
[1] J-H Kim and B Y Lattimer ldquoReal-time probabilistic classifi-cation of fire and smoke using thermal imagery for intelligentfirefighting robotrdquo Fire Safety Journal vol 72 pp 40ndash49 2015
[2] J W Starr and B Y Lattimer ldquoEvaluation of navigation sensorsin fire smoke environmentsrdquo Fire Technology vol 50 no 6 pp1459ndash1481 2014
[3] J-H Kim B Keller and B Y Lattimer ldquoSensor fusion basedseek-and-find fire algorithm for intelligent firefighting robotrdquoin Proceedings of the IEEEASME International Conference onAdvanced Intelligent Mechatronics (AIM rsquo13) pp 1482ndash1486IEEE Wollongong Australia July 2013
[4] J G McNeil J Starr and B Y Lattimer ldquoAutonomous fire sup-pression using multispectral sensorsrdquo in Proceedings of theIEEEASME International Conference on Advanced IntelligentMechatronics Mechatronics for HumanWellbeing (AIM rsquo13) pp1504ndash1509 Wollongong Austrailia July 2013
[5] J-H Kim J W Starr and B Y Lattimer ldquoFirefighting robotstereo infrared vision and radar sensor fusion for imagingthrough smokerdquo Fire Technology vol 51 no 4 pp 823ndash8452015
[6] RC Luo andK L Su ldquoAutonomous fire-detection systemusingadaptive sensory fusion for intelligent security robotrdquo IEEEASME Transactions on Mechatronics vol 12 no 3 pp 274ndash2812007
[7] M A Jackson and I Robins ldquoGas sensing for fire detectionmeasurements of CO CO
2
H2
O2
and smoke density in Euro-pean standard fire testsrdquo Fire Safety Journal vol 22 no 2 pp181ndash205 1994
[8] B U Toreyin R G Cinbis Y Dedeoglu and A E Cetin ldquoFiredetection in infrared video using wavelet analysisrdquo OpticalEngineering vol 46 no 6 Article ID 067204 2007
[9] B U Toreyin Y Dedeoglu andA E Cetin ldquoWavelet based real-time smoke detection in videordquo in Proceedings of the 13th Euro-pean Signal Processing Conference pp 4ndash8 Antalya TurkeySeptember 2005
[10] T Celik H Demirel H Ozkaramanli and M Uyguroglu ldquoFiredetection using statistical color model in video sequencesrdquoJournal of Visual Communication and Image Representation vol18 no 2 pp 176ndash185 2007
[11] L Merino F Caballero J R Martınez-de-Dios I Maza and AOllero ldquoAn unmanned aircraft system for automatic forest firemonitoring and measurementrdquo Journal of Intelligent amp RoboticSystems vol 65 no 1 pp 533ndash548 2012
[12] Y Wang T W Chua R Chang and N T Pham ldquoReal-timesmoke detection using texture and color featuresrdquo in Proceed-ings of the 21st International Conference on Pattern Recognition(ICPR rsquo12) pp 1727ndash1730 Tsukuba Japan November 2012
[13] G Marbach M Loepfe and T Brupbacher ldquoAn image process-ing technique for fire detection in video imagesrdquo Fire SafetyJournal vol 41 no 4 pp 285ndash289 2006
[14] M I Chacon-Murguia and F J Perez-Vargas ldquoThermal videoanalysis for fire detection using shape regularity and intensitysaturation featuresrdquo in Pattern Recognition J F Martınez-Trinidad J A Carrasco-Ochoa C B-Y Brants and E RHancock Eds vol 6718 of Lecture Notes in Computer Sciencepp 118ndash126 Springer Berlin Germany 2011
[15] W Phillips III M Shah and N D V Lobo ldquoFlame recognitionin videordquo Pattern Recognition Letters vol 23 no 1ndash3 pp 319ndash327 2002
[16] D Han and B Lee ldquoDevelopment of early tunnel fire detectionalgorithm using the image processingrdquo in Advances in VisualComputing pp 39ndash48 Springer Berlin Germany 2006
[17] Y Chunyu F Jun W Jinjun and Z Yongming ldquoVideo firesmoke detection using motion and color featuresrdquo Fire Technol-ogy vol 46 no 3 pp 651ndash663 2010
[18] F Yuan ldquoVideo-based smoke detection with histogram se-quence of LBP and LBPV pyramidsrdquo Fire Safety Journal vol 46no 3 pp 132ndash139 2011
[19] F Lafarge X Descombes and J Zerubia ldquoTextural kernel forSVM classification in remote sensing application to forest firedetection andUrban area extractionrdquo in Proceedings of the IEEEInternational Conference on Image Processing (ICIP rsquo05) pp1096ndash1099 September 2005
[20] F Amon and A Ducharme ldquoImage frequency analysis for test-ing of fire service thermal imaging camerasrdquo Fire Technologyvol 45 no 3 pp 313ndash322 2009
[21] F Amon V Benetis J Kim and A Hamins ldquoDevelopmentof a performance evaluation facility for fire fighting thermalimagersrdquo in Defense and Security pp 244ndash252 2004
[22] F D Maxwell ldquoA portable IR system for observing fire thrusmokerdquo Fire Technology vol 7 no 4 pp 321ndash331 1971
[23] A Barducci D Guzzi P Marcoionni and I Pippi ldquoInfrareddetection of active fires and burnt areas theory and observa-tionsrdquo Infrared Physics amp Technology vol 43 no 3ndash5 pp 119ndash125 2002
[24] C Wang and S Qin ldquoAdaptive detection method of infraredsmall target based on target-background separation via robustprincipal component analysisrdquo Infrared Physics amp Technologyvol 69 pp 123ndash135 2015
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
12 Journal of Sensors
In the near future the classification of fire smoke andtheir thermal reflections will be evaluated on any classifiersand features to increase performanceThe convolution neuralnetwork of deep learning which has recently shown highperformance could be explored as a classifier also model-based image features such as discrete wavelet transform willbe further studied
Appendix
See Figure 9
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this manuscript
Acknowledgments
This work was sponsored by the Office of Naval ResearchGrant no N00014-11-1-0074 scientific office Dr ThomasMcKenna in USA Hwarang-dae Research Institute in Seouland Agency for Defense Development in Daejeon SouthKorea The authors would like to thank Mr Joseph Starr andMr JoshMcNeil for assisting in performing the fire testsTheauthors would also like to thank Rosana K Lee for helpingand supporting this research
References
[1] J-H Kim and B Y Lattimer ldquoReal-time probabilistic classifi-cation of fire and smoke using thermal imagery for intelligentfirefighting robotrdquo Fire Safety Journal vol 72 pp 40ndash49 2015
[2] J W Starr and B Y Lattimer ldquoEvaluation of navigation sensorsin fire smoke environmentsrdquo Fire Technology vol 50 no 6 pp1459ndash1481 2014
[3] J-H Kim B Keller and B Y Lattimer ldquoSensor fusion basedseek-and-find fire algorithm for intelligent firefighting robotrdquoin Proceedings of the IEEEASME International Conference onAdvanced Intelligent Mechatronics (AIM rsquo13) pp 1482ndash1486IEEE Wollongong Australia July 2013
[4] J G McNeil J Starr and B Y Lattimer ldquoAutonomous fire sup-pression using multispectral sensorsrdquo in Proceedings of theIEEEASME International Conference on Advanced IntelligentMechatronics Mechatronics for HumanWellbeing (AIM rsquo13) pp1504ndash1509 Wollongong Austrailia July 2013
[5] J-H Kim J W Starr and B Y Lattimer ldquoFirefighting robotstereo infrared vision and radar sensor fusion for imagingthrough smokerdquo Fire Technology vol 51 no 4 pp 823ndash8452015
[6] RC Luo andK L Su ldquoAutonomous fire-detection systemusingadaptive sensory fusion for intelligent security robotrdquo IEEEASME Transactions on Mechatronics vol 12 no 3 pp 274ndash2812007
[7] M A Jackson and I Robins ldquoGas sensing for fire detectionmeasurements of CO CO
2
H2
O2
and smoke density in Euro-pean standard fire testsrdquo Fire Safety Journal vol 22 no 2 pp181ndash205 1994
[8] B U Toreyin R G Cinbis Y Dedeoglu and A E Cetin ldquoFiredetection in infrared video using wavelet analysisrdquo OpticalEngineering vol 46 no 6 Article ID 067204 2007
[9] B U Toreyin Y Dedeoglu andA E Cetin ldquoWavelet based real-time smoke detection in videordquo in Proceedings of the 13th Euro-pean Signal Processing Conference pp 4ndash8 Antalya TurkeySeptember 2005
[10] T Celik H Demirel H Ozkaramanli and M Uyguroglu ldquoFiredetection using statistical color model in video sequencesrdquoJournal of Visual Communication and Image Representation vol18 no 2 pp 176ndash185 2007
[11] L Merino F Caballero J R Martınez-de-Dios I Maza and AOllero ldquoAn unmanned aircraft system for automatic forest firemonitoring and measurementrdquo Journal of Intelligent amp RoboticSystems vol 65 no 1 pp 533ndash548 2012
[12] Y Wang T W Chua R Chang and N T Pham ldquoReal-timesmoke detection using texture and color featuresrdquo in Proceed-ings of the 21st International Conference on Pattern Recognition(ICPR rsquo12) pp 1727ndash1730 Tsukuba Japan November 2012
[13] G Marbach M Loepfe and T Brupbacher ldquoAn image process-ing technique for fire detection in video imagesrdquo Fire SafetyJournal vol 41 no 4 pp 285ndash289 2006
[14] M I Chacon-Murguia and F J Perez-Vargas ldquoThermal videoanalysis for fire detection using shape regularity and intensitysaturation featuresrdquo in Pattern Recognition J F Martınez-Trinidad J A Carrasco-Ochoa C B-Y Brants and E RHancock Eds vol 6718 of Lecture Notes in Computer Sciencepp 118ndash126 Springer Berlin Germany 2011
[15] W Phillips III M Shah and N D V Lobo ldquoFlame recognitionin videordquo Pattern Recognition Letters vol 23 no 1ndash3 pp 319ndash327 2002
[16] D Han and B Lee ldquoDevelopment of early tunnel fire detectionalgorithm using the image processingrdquo in Advances in VisualComputing pp 39ndash48 Springer Berlin Germany 2006
[17] Y Chunyu F Jun W Jinjun and Z Yongming ldquoVideo firesmoke detection using motion and color featuresrdquo Fire Technol-ogy vol 46 no 3 pp 651ndash663 2010
[18] F Yuan ldquoVideo-based smoke detection with histogram se-quence of LBP and LBPV pyramidsrdquo Fire Safety Journal vol 46no 3 pp 132ndash139 2011
[19] F Lafarge X Descombes and J Zerubia ldquoTextural kernel forSVM classification in remote sensing application to forest firedetection andUrban area extractionrdquo in Proceedings of the IEEEInternational Conference on Image Processing (ICIP rsquo05) pp1096ndash1099 September 2005
[20] F Amon and A Ducharme ldquoImage frequency analysis for test-ing of fire service thermal imaging camerasrdquo Fire Technologyvol 45 no 3 pp 313ndash322 2009
[21] F Amon V Benetis J Kim and A Hamins ldquoDevelopmentof a performance evaluation facility for fire fighting thermalimagersrdquo in Defense and Security pp 244ndash252 2004
[22] F D Maxwell ldquoA portable IR system for observing fire thrusmokerdquo Fire Technology vol 7 no 4 pp 321ndash331 1971
[23] A Barducci D Guzzi P Marcoionni and I Pippi ldquoInfrareddetection of active fires and burnt areas theory and observa-tionsrdquo Infrared Physics amp Technology vol 43 no 3ndash5 pp 119ndash125 2002
[24] C Wang and S Qin ldquoAdaptive detection method of infraredsmall target based on target-background separation via robustprincipal component analysisrdquo Infrared Physics amp Technologyvol 69 pp 123ndash135 2015
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
Journal of Sensors 13
[25] N Aggarwal and R K Agrawal ldquoFirst and second order sta-tistics features for classification of magnetic resonance brainimagesrdquo Journal of Signal and Information Processing vol 3 no2 pp 146ndash153 2012
[26] B Ko K-H Cheong and J-Y Nam ldquoEarly fire detectionalgorithm based on irregular patterns of flames and hierarchicalBayesian Networksrdquo Fire Safety Journal vol 45 no 4 pp 262ndash270 2010
[27] H Maruta Y Kato A Nakamura and F Kurokawa ldquoSmokedetection in open areas using its texture features and time seriespropertiesrdquo in Proceedings of the IEEE International Symposiumon Industrial Electronics (ISIE rsquo09) pp 1904ndash1908 IEEE SeoulSouth Korea July 2009
[28] C M Bautista C A Dy M I Manalac R A Orbe and MCordel ldquoConvolutional neural network for vehicle detection inlow resolution traffic videosrdquo in Proceedings of the IEEE Region10 Symposium (TENSYMP ) pp 277ndash281 IEEE Bali IndonesiaMay 2016
[29] H Wang Y Cai X Chen and L Chen ldquoNight-time vehiclesensing in far infrared image with deep learningrdquo Journal ofSensors vol 2016 Article ID 3403451 8 pages 2016
[30] C Shen Z Bai H Cao et al ldquoOptical flow sensorINSmagne-tometer integrated navigation system for MAV in GPS-deniedenvironmentrdquo Journal of Sensors vol 2016 Article ID 610580310 pages 2016
[31] A Bruhn J Weickert and C Schnorr ldquoLucasKanade meetsHornSchunck combining local and global optic flow meth-odsrdquo International Journal of Computer Vision vol 61 no 3 pp1ndash21 2005
[32] A S N Huda and S Taib ldquoSuitable features selection formonitoring thermal condition of electrical equipment usinginfrared thermographyrdquo Infrared Physics andTechnology vol 61pp 184ndash191 2013
[33] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[34] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975
[35] P Harrington Machine Learning in Action Manning Publica-tions 2012
[36] D J Hand HMannila and P Smyth Principles of DataMiningMIT Press 2001
[37] D Lin X Xu and F Pu ldquoBayesian information criterionbased feature filtering for the fusion of multiple features inhigh-spatial-resolution satellite scene classificationrdquo Journal ofSensors vol 2015 Article ID 142612 10 pages 2015
[38] F Van Der Heijden R Duin D De Ridder and D M TaxClassification Parameter Estimation and State Estimation AnEngineering Approach Using MATLAB John Wiley amp Sons2005
[39] R Kohavi ldquoA study of cross-validation and bootstrap foraccuracy estimation and model selectionrdquo in Proceedings ofthe 14th International Joint Conference on Artificial Intelligence(IJCAI rsquo95) Montreal Canada August 1995
[40] K Deb Multi-Objective Optimization Using Evolutionary Algo-rithms vol 16 John Wiley amp Sons New York NY USA 2001
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of
International Journal of
AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation httpwwwhindawicom
Journal ofEngineeringVolume 2014
Submit your manuscripts athttpwwwhindawicom
VLSI Design
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Shock and Vibration
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation httpwwwhindawicom
Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
DistributedSensor Networks
International Journal of