research article an obstacle recognizing mechanism for...

Research ArticleAn Obstacle Recognizing Mechanism for AutonomousUnderwater Vehicles Powered by Fuzzy Domain Ontology andSupport Vector Machine

Zhen-Shu Mi Ahmad C Bukhari and Yong-Gi Kim

Department of Computer Science and Engineering Research Institute Gyeongsang National UniversityJinju Kyungnam 660-701 Republic of Korea

Correspondence should be addressed to Yong-Gi Kim ygkimgnuackr

Received 12 December 2013 Revised 11 June 2014 Accepted 18 July 2014 Published 5 August 2014

Academic Editor Tsung-Chih Lin

Copyright copy 2014 Zhen-Shu Mi et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

The autonomous underwater vehicle (AUV) and the problems associated with its safe navigation have been studied for the last twodecades The real-time underwater obstacle recognition procedure still has many complications associated with it and the issuebecomes worse with vague sensor data These problems can be coped with the merger of a robust classification mechanism and adomain knowledge acquisition technique In this paper we introduce a hybrid mechanism to recognize underwater obstacles forAUV based on fuzzy domain ontology and support vector machine (SVM) SVM is an efficient algorithm that was developed forrecognizing 3Dobject in recent years and is a new generation learning systembased on recent advances in statistical learning theoryThe amalgamation of fuzzy domain ontology with SVM boosts the performance of the obstacle recognition module by providingthe timely semantic domain information of the surrounding circumstances Also the reasoning ability of the fuzzy domain ontologycan expedite the obstacle avoidance process In order to evaluate the performance of the system we developed a prototype simulatorbased on OpenGL and VC++ We compared the outcomes of our proposed technique with backpropagation algorithm and classicSVM based techniques

1 Introduction and Background

Currently the underwater obstacle recognition issue hasreceived a lot of attention and due to its significance hasbecome one of the hot research topics of machine learningand computer vision It has been observed that appearance-based object recognition methods can perform better in avariety of ongoing problems On the other hand many ofthese methods either require good whole-object segmenta-tion which severely limits their performance in the presenceof clutter occlusion or background changes or utilize simpleconjunctions of low-level features which cause crosstalkproblems as the number of objects is increased [1] Inthis research we investigated an appearance-based obstaclerecognition mechanism using fuzzy ontology and SVM thatsolves many of these problems for AUVs navigation safety

SVM is a new generation learning system based onrecent advances in statistical learning theory SVMs deliver

state-of-the-art performance in real-world applications suchtext categorization hand-written character recognition andimage classificationTheir first introduction in the early 1990sled to a recent explosion of applications and the deepeningof theoretical analysis has now established support vectormachines along with neural networks as one of the standardtools for machine learning and data mining It is consideredas a very effectivemethod of learning and classification [2ndash4]

Because of their commercial and military potentialand the technological challenges in developing themautonomous underwater vehicles (AUVs) have becomea hot topic of research in ocean sciences There are twomain reasons to keep navigation security for AUVs Firstsafe AUVs can search rescue and facilitate explorationin the scientific field Second it can save a lot of moneybecause of the expensive cost of AUVs Therefore the AUVssafety problem has become an exciting hotspot researchtopic Nevertheless the multifarious nature of underwater

Hindawi Publishing CorporationMathematical Problems in EngineeringVolume 2014 Article ID 676729 10 pageshttpdxdoiorg1011552014676729

2 Mathematical Problems in Engineering

environments makes the research work a difficult task tohandle Thereby development of an intelligent 3D obstaclerecognition system becomes necessary for AUVs safenavigation Some notable researches recently have beenexecuted for obstacle detection of AUVs using SVM andother techniques [5ndash9]

In a normal situation the AUVs use sonar instead ofa good quality camera to recognize obstacles AUV cannotsee even several meters ahead since underwater situationis very dark in the deep ocean or the water is usually notclear in shallow water However in some special cases highdefinition cameras are used to achieve the desired resultsFor instance when a vessel needs to be cleaned or repairedfor some malfunction a good quality camera is a very highpriority choice to solve such kind of problems There are alot of applications for taking some valuable pictures usingcameras including underwater resource exploration bridgesurvey and commercial diving

In this research we introduce a fast underwater obstaclerecognition system for AUV based on fuzzy ontology andSVM [3 10 11] The aim of this paper is to prove that themerger of fuzzy ontology and SVM demonstrates better per-formance in comparison with BP algorithm and simple SVMtechnique The dynamic 3D obstacle simulation designed byOpenGL will provide some test data for getting the betterexperiment results Therefore the SVM will help underwa-ter obstacle recognition mission planning and intelligentcontrol of the AUV in a variety of ways [12ndash14] At presentontology has proved itself as an effective technology forrelevant information extraction and knowledge sharing Weused the ontology to store and exploit the domain knowledgeof the underwater water domain This unique power of theontology can be exploited to design an accurate knowledgerepresentation system

During the past few years researchers are using ontologywidely in different domains for knowledge representation andinformation extraction It is another fact that most of theinformation resides in offline and online resources and is inimprecise format The ontology only deals with crisp dataand cannot find desired results from vague sources of data[15 16] To address this issue researchers have incorporatedfuzzy logic theory into classic ontology Fuzzy theory iswidelyknown among the research community due to its ability todeal with blur information [17 18] The rest of the paperis organized as follows Section 2 describes the conceptsand working of SVM however Section 3 highlights theimportance of fuzzy ontology and its integration with SVMThe fourth part of this paper presents the proposed schemeSections 5 and 6 belong to experiments and conclusionsrespectively

2 The Support Vector Machine

The foundation of support vector machine (SVM) was devel-oped by Vapnik in 1995 [12] SVM is a powerful classificationmethodology that can create specialized functions from a setof labeled training dataThese functions can be a classificationfunction or a general regression function For classification

SVM operates by finding a hypersurface in the space ofpossible inputs This hypersurface will attempt to split thepositive examples from the negative examples The split willbe chosen to have the largest distance from the hypersurfaceto the nearest of the positive and negative examples Intu-itively this makes the classification correct for testing datathat is near but not identical to the training dataThe goal ofSVM is to find out an optimally separable hyperplane to solvethe classification task The optimal hyperplane has maximaldistance between the two classes and the hyperplane

21 Linear Separable Case Consider the problem of sepa-rating the set of training vectors belonging to two linearlyseparable classes

(119909119894 119910119894) 119909 isin 119877119889 119910 isin +1 minus1 119894 = 1 119899 (1)

with a hyperplane

119908 sdot 119909 + 119887 = 0 (2)

where the parameters 119908 119887 are constrained by

min119894

1003816100381610038161003816119908 sdot 119909119894 + 1198871003816100381610038161003816 = 1 (3)

The set of vectors is said to be optimally separated bya hyperplane if it is separated without error The distancebetween the closest vector and the hyperplane is maximalHence a separating hyperplane in canonical form mustsatisfy the following constraints

119910119894sdot (119908 sdot 119909

119894+ 119887) ge 1 119894 = 1 2 119899 (4)

Suppose that the following bound holds

1199082 le 119888 (5)

The VC dimension ℎ of the set of canonical hyperplanesin 119899 dimensional space is bounded by

ℎ le min ([1198772 119888] 119889) + 1 (6)

where 119877 is the radius of a hypersphere enclosing all trainingvectors Hence the hyperplane that optimally separates thedata is the one that minimizes (4)

120601 (119908) =1

2(119908 sdot 119908) (7)

This is equivalent to minimizing an upper bound on theVC dimension In practice such a hyperplane does not existSo we need to relax the constraints of (4) by introducing slackvariables 120585 ge 0 119894 = 1 2 119899

119910119894sdot (119908 sdot 119909

119894+ 119887) ge 1 minus 120585 119894 = 1 2 119899 (8)

In this case the optimization problem becomes

120601 (119908 120585) =1

2(119908 sdot 119908) + 119862

119899

sum119894=1

120585119894 (9)

Mathematical Problems in Engineering 3

with a user defined positive finite constant 119862 The solution tooptimization problem (9) under the constraints of (8) couldbe obtained in the saddle point of Lagrangian function

119871 (119908 119887 120572 120585 120574) =1

2(119908 sdot 119908) + 119862

119899

sum119894=1

120585119894

minus119899

sum119894=1

120572119894

1003816100381610038161003816119910119894 (119908 sdot 119909119894 + 119887) minus 1 + 1205851198941003816100381610038161003816 minus119899

sum119894=1

120574119894120585119894

(10)

where 120572119894ge 0 120574 ge 0 and 119894 = 1 2 119899 are the Lagrange

multipliersThe Lagrangian function has to be minimized with

respect to 119908 119887 and 120585119894 Classical Lagrangian duality enables

the primal problem (10) to be transformed to its dualproblem which is easier to solve The dual problem is givenby

max120572

[

[

119899

sum119894=1

120572119894minus1

2

119899

sum119894119895=1

120572119894120572119895119910119894119910119895(119909119894 119909119895)]

]

(11)

with constraints119899

sum119894=1

120572119894119910119894= 0 (12)

0 le 120572119894le 119862 119894 = 1 2 119899 (13)

This is a classic quadratic optimization problem Thereexists a unique solution According to the Kuhn-Tuckertheorem of optimization theory [8] the optimal solutionsatisfies

120572119894[119910119894(119908 sdot 119909

119894+ 119887) minus 1] = 0 119894 = 1 2 119899 (14)

Equation (14) has nonzero Lagrangemultiplierswhen andonly when the points 119909 satisfy

119910119894sdot (119908 sdot 119909

119894+ 119887) = 1 (15)

These points are termed support vectors (SV)The hyper-plane is determined by the SV which is a small subset of thetraining vectors Hence if 120572lowast

119894is the nonzero optimal solution

the classifier function can be expressed as

119891 (119909) = sgn119899

sum119894=1

120572lowast119894119910119894(119909119894sdot 119909) + 119887lowast (16)

where 119887lowast is the solution of (14) for any the nonzero 120572lowast119894

22 Nonlinear Case When a linear boundary is inappropri-ate SVM can map the input vector into a high dimensionalfeature space By defining a nonlinear mapping the SVMconstructs an optimal separating hyperplane in this higherdimensional space Usually nonlinear mapping is defined asfollowing the function

120601 (sdot) 119877119889 997888rarr 119877119889ℎ (17)

In this case optimal function (8) becomes (18) with thesame constraints

max120572

[

[

119899

sum119894=1

120572119894minus1

2

119899

sum119894119895=1

120572119894120572119895119910119894119910119895119870(119909119894 119909119895)]

]

(18)

where

119870(119909119894 119909119895) = 120601 (119909

119894) sdot 120601 (119909

119895) (19)

is the kernel function performing the nonlinearmapping intofeature space It may be any of the symmetric functions thatsatisfy theMercer conditions [14]The following functions arecommonly used kernel functions

(1) Polynomial function

119870(119909119894 119909119895) = (119909

119894sdot 119909119895+ 1)119902

119902 gt 0 (20)

(2) Radial basis function

119870(119909119894 119909119895) = exp

minus

10038161003816100381610038161003816119909119894 minus 119909119895100381610038161003816100381610038162

1205902

(21)

(3) Sigmoid function

119870(119909119894 119909119895) = tanh (120573 sdot (119909

119894sdot 119909119895) + 119888) (22)

The classifier function can be defined as

119891 (119909) = sgn119899

sum119894=1

120572lowast119894119910119894119870(119909119894 119909) + 119887lowast (23)

where 119887lowast is the solution of the following equation for any ofthe nonzero 120572lowast

119894

120572119894[

[

119910119894

119899

sum119895=1

120572119895119910119895119870(119909119895 119909119894) + 119887 minus 1]

]

= 0 (24)

3 The Fuzzy Domain Ontology

Ontology is a branch of metaphysics which focuses on thestudy of existence Ontology is an explicit formal specifica-tion of a shared conceptualization of any domain which ismachine readable and human understandable format [15]In ontology we basically focus on concepts (classes) of thedomain their values and their properties Ontology arrangesthe classes in a hierarchical structure in the form of sub-classes and superclasses hierarchies and also defines whichproperty has constraints on its values Basically ontology isdeveloped to share a common understanding of the domainknowledge among people and software in order to reuse theclasses of a domain instead of remodelling them Ontologyis written in a specific language called OWL (Web OntologyLanguage) [16 17] which is developed by W3C OWL isspecially designed for ontology modelling Figure 1 depictsthe graphical model fuzzy underwater ontology generated by


Figure 1 Graphical model of fuzzy underwater environment ontology

a Protege OnToGraf tool OWL-2 is the extended version ofOWL which holds new features and rationale In order toincrease the expressivity level ofOWL several new constructswere introduced along with their unique modelling featuressuch as extended annotations and complex data representa-tions Ontology can be expressed in equation format as

Ontology

= (Concepts CProperties PRelationalships R

Values VConstraints of property value VC)

(25)

In the above expression notations C P R V andVC represent concepts properties of concepts relationshipamong concepts values of the concepts and constraints onproperty values respectively The Ontology can be classifiedinto metaontology domain ontology and application levelontology Usually we use domain ontology which is used torepresent the information about a specific domain On theother hand there is no exactly one way to define ontology[15 17] Researchers used differentways to define the ontologyaccording to their needs and demandsThe values of conceptsand properties in the classical domain ontology are crispbut as mentioned earlier most real-time systems nowadaysare based on complex structure which is based on fuzzy settheory

31 Development of Fuzzy Domain Ontology In order tomake the paper self-contained we first define the conceptdefinitions and terminologies before moving towards theformal development of underwater fuzzy ontology Fuzzy settheory was introduced by Lotfi Zadeh in 1965 [15 18] to dealwith vague and imprecise concepts In classical set theoryelements either belong to a particular set or do not Thereis no concept of partial membership in classical set theory

However in fuzzy set theory the association of an elementwith a particular set lies between 0 and 1 which is called itsdegree of association ormembership degree Fuzzy set theoryadds generalization concept in to classical set theory andmakes it diverse enough to represent imprecise boundarieslike hot tall low speed and so forth A fuzzy set ldquoFSrdquo over theuniverse of discourse ldquo119911rdquo can be defined by its membershipfunction 120583FS whichmaps element ldquo119911rdquo to values between [0 1]

120583FS (119911) 119911 997888rarr [0 1] (26)

Here 119911 isin 119885 and 120583FS(119911) provide the degree of membershipbywhich 119911 belongs to119885 119911 is considered as fullmember of119885 if120583FS(119910) = 1 and is considered partial if 120583FS(119910) is between 0 and1 say 057 The strength of our proposed system is the fuzzyenriched domain ontology which not only is able to presentthe associated domain knowledge of any obstacle but canalso assist obstacle avoidance and intelligent path planningby providing timely information about obstacles

To achieve perfection in the results and to expeditethe process of fuzzy based domain ontology we employedthe service of the domain experts The domain experts donot have knowledge about SVM and ontology schemes butthey can best describe the underwater environment andobstacles The domain experts monitored all the phases ofthe ontology development Furthermore we followed theontology development steps which are defined in [19] todevelop the underwater domain ontology The steps can bedescribed as follows

(1) Determine the domain and scope of the ontology(2) Consider reusing existing ontologies(3) Enumerate important terms in the ontology(4) Define the classes and the class hierarchy(5) Define the properties of classes


FC-N r1 r2 r3 r4 rn

Fc1P3Fc1P1Fc1P2

Fc1Pn

Fc2P3Fc2P1Fc2P2

Fc2Pn

Fc3P3Fc3P1Fc3P2

Fc3Pn

FcnP3FcnP1FcnP2

FcnPn

F1s1)

F1s3)F1s2)

F1sk)

F1sn)

F2s1)F2s2)

F2s3)F2sk)

F2sn)

F3s1)F3s2)

F3s3)F3sk)

F3sn)

Fns1)

Fns2)

Fns3)

Fnsk

Fnsn)

Main category layer

Fuzzy class and relationship layer

Fuzzy class property layer

Fuzzy sets layer

CAT-1 CAT-2 CAT-3 CAT-N

FC-1 FC-2 FC-3

middot middot middot middot middot middotmiddot middot middot

middot middot middot





Figure 2 Structure of fuzzy underwater domain ontology

Figure 3 Images of 3 objects of the COIL database

(6) Define the facets of the slots(7) Create instances

To author the fuzzy based domain ontology we chose theProtege 41 and OWL-2 due to their technical soundness ininformation engineering circle Figure 2 displays the architec-ture of fuzzy domain ontology

4 The Proposed 3D ObstacleRecognition System

The SVM in a linear separable case categorizes the same classof data on one side of the optimal hyperplane and the oppositeclass on the other sideWe illustrate an example to explain thisproblem see Figure 1 We took a set of data from the COILimage database for virtual experiment purposes The COIL(Columbia Object Image Library) database consists of 7200images of 100 objects We selected some simple and typicalimages like simulation obstacles by way of example test data(eg Figure 3) The images are color images (24 bits for eachof the RGB channels) of 128 times 128 pixels For each object theturntable is rotated 5 degrees per imageThe tester can designand adjust the rotation degree to get an optimal test image

data All the selected color image data are according to COILimage standard Figures 1 and 4 show a selection of the objectsin the database and one every three views of a specific objectrespectively [20]

41 Processing Based onOpenGL Simulation In real researchthe underwater environment is hazy and surrounding infor-mation is not known However it is easy to understand thesimulation using OpenGL and we placed some obstaclesrandomly and created the virtual environment according tothe real system For the experimental purpose we designeda 3D virtual testbed using OpenGL to get some test datato prove the effectiveness of our proposed algorithm Eachimage I = (R G B) was first transformed into a gray-levelimage through the conversion formula

Gray = Red lowast 30 + Green lowast 59 + Blue lowast 11 (27)

rescaling the obtained gray-level from 0 to 255The obstaclesimages spatial resolution was reduced to 32 times 32 by averagingthe gray levels over 4 times 4 pixel patches The purpose of thispreprocessing is to identify the selected data for the imagecontaining obstacles and standardize the obstacles images[20 21]

42 Training and Testing the SVM with Semantic DomainData It is a very common problem to select kernel functionsin the SVM training phase In our experiments the threekernel parameters 120590 119888 and 119902 determine the dynamicsof an SVM network We selected the optimal parametersfor optimizing and adjusting the SVM parameters Theradial basis function (RBF) kernel function nonlinearly maps


Figure 4 36 images of all the 72 images of one COIL object

Object

Nonobject

Figure 5 SVM feature space for object recognition based on fuzzy domain ontology

samples into a higher dimensional spaceThe RBF kernel hasless numerical difficulties its computation is not as complexas the other kernels Thus the RBF kernel is a better choicefor training and testing We used semantically annotatedimages to train our SVM classifier In the SVM recognitionsystem we adjusted the kernel parameters by training untilthe optimal kernel parameters were found To reduce thecomputational quantity we used the principal componentanalysis method to generate 128-dimensional feature vectorswhich can represent the characteristics of the image [12]

In one case 20 objects total 720 images are selected assamples in the training phase All the 1440 images for 20objects are the test data Given a subset of the 20 objects andthe associated training set of 36 images for each object thetest set consists of the remaining 36 images for each objectFirst we tested the recognition system on sets of 20 of the100 COIL objects The training sets consisted of 36 images

(every 10 degrees for each of the 20 objects) and the test setsconsisted of the remaining 36 images For all the 12 randomchoices of 20 of the 100 objects the system got a perfect scoreSo we decided to select by hand the 20 objects that weremore difficult to recognize To gain a better understandingof how an SVM actually performs recognition it may beuseful to look at the relative weights of the components of 119908A gray-values encoded representation of the absolute valueof the components of 119908 relative to the optimal separatinghyperplane [20] Note that the background is essentiallyirrelevant while larger components can be found in thecentral portion of the image [12 20] Figure 5 shows the SVMfeature space based on fuzzy domain ontology

43 Proposed System Overview The proposed system con-sists of five functional phases named as the detection andacquisition phase obstacle analysis phase initial recognition


Detection and acquisition phase

Datacapturing

Dataacquisition

Datatransmitting

Obstacle analysis phase

Preprocessing Imageprocessing

Featureextraction

Initial recognition phase

Image training and testing SVM classificationbased on ontology data

Categorical recognition phase

Ontologyfeature

mappingOntologyreasoning

Hierarchicalobstacle

categorization

Decision taking phase

Obstacle avoidance Path planning

Figure 6 The SVM and fuzzy ontology supported obstacle recognition system architecture

phase categorical recognition phase and decision makingphase While maneuvering the AUV continuously takesimages of its surroundings and sends them back to thecoastal station for analysis The sophisticated analysis phaseapplies the predefined algorithms against the image data toget results Subsequently these analysis results are forwardedto the decisionmakingmodule which is specially designed tomake intelligent decisions and sends the instructions back tothe AUV for safe navigation In the subheadings we elaboratethe system in detail

431 Detection and Acquisition Phase This phase can befurther divided into three phases which are data capturingdata acquisition and data transmitting To accomplish theunderwater mission an AUV has to acquire optimal pathinformation from the onboard system or from the surfacecontrol station To the get the advice for safemaneuvering theAUV repeatedly captures the images with an onboard highdefinition camera Usually the AUV has internal memorywhich stores the current situation and the transmissionfunction of the detection and acquisition phase assists insending the image data back to the surface station for theanalysis

432 Obstacle Analysis Phase Similar to the previous phasethis phase can also be divided into a series of subactivities Inthe preprocessing stage the system acquires the image dataand performs the initial screening The initially processedimages are automatically sent to the image processing phasewhich is the core module The job of this module is to deeplyanalyze the images after sharpening and noise removingFurther from an image it can extract the features such ascolor width height diameter shape texture and orientationtowards the AUV

433 Initial Recognition Phase This phase is consideredthe most important phase of the entire system In factthe efficiency of the entire system is dependent upon thisphase This phase performs two main tasks It calls the SVMclassifier and starts the training and testing activities Beforethe testing we have to train the classifier with sample data Inour technique we introduced a fuzzy ontology based trainerwhich trains the SVM classical algorithm with semanticallyannotated image dataThis allows the classifier to be aware ofthe color texture geometry and orientation of the image Asthe SVM is a binary classifier at initial phase it segregates theobstacle from nonobstacle data as shown in Figure 6


434 Categorical Recognition Phase The categorical recog-nition phase is based on the fuzzy ontology scheme Themain purpose of this phase is to categorize the detectedobstacles into classes and subclasses We introduced severalfuzzy classes and subclasses in our ontology as the followingunderwater animals rock icebergs long obstacle shortobstacle dangerous obstacle very dangerous obstacle softobstacle and hard obstacle Each class has its own propertiesand values The image processing phase extracts the obstaclefeatures and categorical recognition phase compares theirproperties with semantically defined properties and segre-gates the obstacles according to their matching classes Thiscategorical recognition is essential to expedite the processof decision making For instance if the fuzzy ontologicalclassifier finds an obstacle as soft or nondangerous it simplyadvises the AUV not to change the route We use a Pelletreasoner in Protege to reason against the supplied data [1516]

435 Decision Making Phase Last but not least all theanalysis results are forwarded to the decision making phaseThe underlying algorithm performs the computations andfinds an appropriate decision Lastly the commands aretransmitted to the AUV Obstacle avoidance and path plan-ning do not fall in the scope of our paper Figure 6 describesthe proposed system architecture

5 Experiments and Results

To evaluate the efficiency of our proposed system we devel-oped a virtual testbed which holds all the components of thereal-world environment In Figure 7 we considered the smallblack ball as a starting point and the right red ball as an endpoint In our algorithmwe added underwater survey as a taskfor the AUV When we play the simulation the AUV startsgrabbing the surroundings information and takes the highdefinition images with the help of the mounted camera Aftercapturing the images from the virtual environment the AUVsends those to the base station for further processing [21]

The base station computer gets the random images as testdata each secondThe simulation generates different kinds ofobstacles around AUVs and facilitates the evaluator to get the3D obstacle images data directly from the experiment Oneof the experiments is tested by a MATLAB tool based on theOpenGL test images [22] In our experiment the OpenGLtest images contain 720 images of 20 objects For each objectwe can generate 36 images from OpenGL simulation andthe simulation is powerful enough to produce a new imageof the same obstacle for every 10 degrees of rotation Weused the same training and testing data sets for SVM andBP algorithms therefore we can get a clear analysis based onboth algorithms

We performed a series of experiments using differenttraining sets to evaluate the performance of the whole system[22 23] For the BP algorithm experiment we selected theoptimal parameters in the experiments while optimizing theneural network architectures To train the artificial neuralnetworks we used the back propagation algorithm We used

Figure 7 The virtual testbed to evaluate the proposed system

Table 1 SVM optimal parameters

Arg 119862 Training error Training time1 100 00106 4809 s1 200 00127 4145 s2 400 00171 3952 s2 500 00062 6315 s

Table 2 Adjusting BP network

Neurons in hidden layer Training error Training time6 00148 63117 s7 00136 68352 s8 00127 49576 s9 00112 45123 s

the tensing function in the hidden layer and logsig functionin the output layer It was trained to find out the optimalneurons which can generate the best output results [4 24]

Tables 1 and 2 have the optimal SVM parameters andadjusted BP network respectively Table 3 shows the perfor-mance results of classic SVM BP and fuzzy ontology basedSVM schemes Figure 8 depicts the average performance ofall three schemes in graphical format We can easily under-stand from the graph that high accuracy rate can be achievedby applying the proposed hybrid approach By applying theproposed scheme not only canwe reduce training and testingtime but we can also improve the efficiency of the systemoverall

6 Conclusion

In this research we introduced a hybridmechanism based onfuzzy domain ontology and support vector machine to expe-dite the process of underwater obstacle recognition Several


Table 3 The performance result of SVMs BP algorithm and the proposed scheme

Obstacles classifier(object sample)

BP trainingtime (s)

SVMtrainingtime (s)

Proposedsystem

training time

BP testingtime (s)

SVM testingtime (s)

Proposedsystem

testing time

BPaccuracy

SVMaccuracy

Proposedsystemaccuracy

5ndash24 7286 1529 1043 028 019 009 8617 9087 979110ndash24 15376 1993 947 391 228 014 8512 8673 981720ndash24 21791 6735 3352 1283 612 126 8494 8990 95775ndash36 14126 4782 2651 293 106 007 8795 9178 983110ndash36 24826 7811 5009 1196 718 233 8621 9299 9747

0

20

40

60

80

100

Training Testing Accuracy

Ann(BP)Classic SVM

Proposed scheme

Figure 8 An average performance comparison graph among BPalgorithm classical SVM and proposed scheme

researchers have addressed this issue and many of them pro-posed solutions to solve this problemHowever the efficiencyof their proposed algorithms decreases with the increase indata vagueness collected from sensors We coupled fuzzyontology with SVM which holds the underwater domaindata in machine readable and human understandable formatwhich further assists the SVM classifier in fast recognitionof the obstacle Moreover the reasoning power of the fuzzyontology provides accurate and timely information to theobstacle avoidance and path planningmodules andmakes theunderwater navigation safe To analyze the performance ofour proposed system we developed a prototype system andperformed a number of experiments by using the training setfrom the COIL database Tomonitor the overall performanceof the system we compared our proposed scheme with BPand classic SVM schemes Our proposed scheme showedmore accurate results in a shorter span of time As a futurework we will train our algorithm with diverse training setsso it can be able to work with any kind of underwater data

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgment

This work was supported by the Korea National ResearchFoundation (NRF) Grant funded by the Korean Government(no 2012R1A1A2038601)

References

[1] T Sergios and K Konstantinos Pattern Recognition ElsevierScience Publishing 2003

[2] C J C Burge ldquoA tutorial on support vector for patternrecognitionrdquo Neural Network pp 1076ndash1121 2001

[3] C Colin ldquoAlgorithmic approaches to training support vectormachines a surveyrdquo in Proceedings of the 8th European Sympo-sium on Artificial Neural Networks pp 27ndash54 2000

[4] E Zheng P Li and Z Song ldquoPerformance analysis andcomparison of neural networks and support vector machinesclassifierrdquo in Proceedings of the 5thWorld Congress on IntelligentControl and Automation pp 52ndash58 June 2004

[5] I Quidu A Hetet Y Dupas and S Lefevre ldquoAUV (Redermor)obstacle detection and avoidance experimental evaluationrdquo inProceedings of OCEANSmdashEurope 2007

[6] S Zhao T Lu and A Anvar ldquoAutomatic object detectionfor AUV navigation using imaging sonar within confinedenvironmentsrdquo in Proceedings of the 4th IEEE Conference onIndustrial Electronics and Applications (ICIEA rsquo09) pp 3648ndash3653 May 2009

[7] Q Li J Zhao and Y Zhao ldquoDetection of ventricular fibril-lation by support vector machine algorithmrdquo in Proceedingsof the International Asia Conference on Informatics in ControlAutomation and Robotics (CAR 09) pp 287ndash290 BangkokThailand February 2009

[8] M S Bewley B Douillard N Nourani-Vatani A FriedmanO Pizarro and S B Williams ldquoAutomated species detectionan experimental approach to kelp detection from sea-floorAUV imagesrdquo in Proceedings of the Australasian Conference onRobotics and Automation (ACRA rsquo12) December 2012

[9] S Karabchevsky B Braginsky and H Guterman ldquoAUV real-time acoustic vertical plane obstacle detection and avoidancerdquoin Proceedings of Autonomous Underwater Vehicles (AUV rsquo12)IEEEOES 2012

[10] V Blanz B Scholkopf H Bulthoff C Burges V Vapnikand V Vetter ldquoComparison of view-based object recognitionalgorithms using realistic 3D modelsrdquo in Proceedings of Inter-national Conference on Artificial Neural Networks vol 112 ofLectureNotes in Computer Science Springer pp 251ndash256 BerlinGermany 1996


[11] H Murase and S K Nayar ldquoVisual learning and recognition of3-d objects from appearancerdquo International Journal of ComputerVision vol 14 no 1 pp 5ndash24 1995

[12] V Vapnik The Nature of Statistical Learning Theory SpringerNew York NY USA 1995

[13] D Bertsekas Nonlinear Programming Athena Scientific Bel-mont Mass USA 1995

[14] R Courant andDHilbertMethods ofmathematical physics vol1 Wildy-Interscience New York NY USA 1953

[15] A C Bukhari and Y-G Kim ldquoIntegration of a secure type-2 fuzzy ontology with a multi-agent platform a proposalto automate the personalized flight ticket booking domainrdquoInformation Sciences vol 198 pp 24ndash47 2012

[16] A C Bukhari and Y Kim ldquoOntology-assisted automatic preciseinformation extractor for visually impaired inhabitantsrdquo Artifi-cial Intelligence Review vol 38 no 1 pp 9ndash24 2012

[17] A C Bukhari and Y Kim ldquoExploiting the heavyweight ontol-ogy with multi-agent system using vocal command system acase study on e-mallrdquo International Journal of Advancements inComputing Technology vol 3 no 6 pp 233ndash241 2011

[18] C Ahmad and Y Kim ldquoIncorporation of fuzzy theory withheavyweight ontology and its application on vague informationretrieval for decision makingrdquo International Journal of FuzzyLogic and Intelligent System vol 11 no 3 pp 171ndash177 2011

[19] A C Bukhari A A Fareedi and Y G Kim ldquoHO2IEV heavy-weight ontology based web information extraction techniquefor visionless usersrdquo in Proceedings of the 7th InternationalConference onNetworked Computing andAdvanced InformationManagement (NCM rsquo11) pp 90ndash95 June 2011

[20] M Pontil and A Verri ldquoSupport vector machines for 3Dobject recognitionrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 20 no 6 pp 637ndash646 1998

[21] A Arif L Young-il J Hee andY Kim ldquoUnsupervised real-timeobstacle avoidance technique based on a hybrid fuzzy methodfor AUVsrdquo International Journal of Fuzzy Logic and IntelligentSystems vol 8 pp 82ndash87 2008

[22] N Yoshikawa and Y Ii ldquoThree-dimensional object recognitionusing multiplex complex amplitude information with supportfunctionrdquo in Proceedings of the 1st International Conference onInnovative Computing Information and Control (ICICIC rsquo06)vol 1 pp 314ndash317 September 2006

[23] J LichengApplication and Realization of Neural Networks XianElectron and Science University 1993

[24] C H Chen ldquoA comparison of neural network models forpattern recognitionrdquo in Proceedings of the 10th InternationalConference on Pattern Recognition pp 45ndash46 June 1990

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MathematicsJournal of


Mathematical Problems in Engineering

Hindawi Publishing Corporationhttpwwwhindawicom

Differential EquationsInternational Journal of

Volume 2014

Applied MathematicsJournal of


Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of


Mathematical PhysicsAdvances in

Complex AnalysisJournal of


OptimizationJournal of


CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of


Operations ResearchAdvances in

Journal of


Function Spaces

Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of Mathematics and Mathematical Sciences


The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Algebra

Discrete Dynamics in Nature and Society



Decision SciencesAdvances in

Discrete MathematicsJournal of


Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Stochastic AnalysisInternational Journal of


environments makes the research work a difficult task tohandle Thereby development of an intelligent 3D obstaclerecognition system becomes necessary for AUVs safenavigation Some notable researches recently have beenexecuted for obstacle detection of AUVs using SVM andother techniques [5ndash9]

In a normal situation the AUVs use sonar instead ofa good quality camera to recognize obstacles AUV cannotsee even several meters ahead since underwater situationis very dark in the deep ocean or the water is usually notclear in shallow water However in some special cases highdefinition cameras are used to achieve the desired resultsFor instance when a vessel needs to be cleaned or repairedfor some malfunction a good quality camera is a very highpriority choice to solve such kind of problems There are alot of applications for taking some valuable pictures usingcameras including underwater resource exploration bridgesurvey and commercial diving

In this research we introduce a fast underwater obstaclerecognition system for AUV based on fuzzy ontology andSVM [3 10 11] The aim of this paper is to prove that themerger of fuzzy ontology and SVM demonstrates better per-formance in comparison with BP algorithm and simple SVMtechnique The dynamic 3D obstacle simulation designed byOpenGL will provide some test data for getting the betterexperiment results Therefore the SVM will help underwa-ter obstacle recognition mission planning and intelligentcontrol of the AUV in a variety of ways [12ndash14] At presentontology has proved itself as an effective technology forrelevant information extraction and knowledge sharing Weused the ontology to store and exploit the domain knowledgeof the underwater water domain This unique power of theontology can be exploited to design an accurate knowledgerepresentation system

During the past few years researchers are using ontologywidely in different domains for knowledge representation andinformation extraction It is another fact that most of theinformation resides in offline and online resources and is inimprecise format The ontology only deals with crisp dataand cannot find desired results from vague sources of data[15 16] To address this issue researchers have incorporatedfuzzy logic theory into classic ontology Fuzzy theory iswidelyknown among the research community due to its ability todeal with blur information [17 18] The rest of the paperis organized as follows Section 2 describes the conceptsand working of SVM however Section 3 highlights theimportance of fuzzy ontology and its integration with SVMThe fourth part of this paper presents the proposed schemeSections 5 and 6 belong to experiments and conclusionsrespectively

2 The Support Vector Machine

The foundation of support vector machine (SVM) was devel-oped by Vapnik in 1995 [12] SVM is a powerful classificationmethodology that can create specialized functions from a setof labeled training dataThese functions can be a classificationfunction or a general regression function For classification

SVM operates by finding a hypersurface in the space ofpossible inputs This hypersurface will attempt to split thepositive examples from the negative examples The split willbe chosen to have the largest distance from the hypersurfaceto the nearest of the positive and negative examples Intu-itively this makes the classification correct for testing datathat is near but not identical to the training dataThe goal ofSVM is to find out an optimally separable hyperplane to solvethe classification task The optimal hyperplane has maximaldistance between the two classes and the hyperplane

21 Linear Separable Case Consider the problem of sepa-rating the set of training vectors belonging to two linearlyseparable classes

(119909119894 119910119894) 119909 isin 119877119889 119910 isin +1 minus1 119894 = 1 119899 (1)

with a hyperplane

119908 sdot 119909 + 119887 = 0 (2)

where the parameters 119908 119887 are constrained by

min119894

1003816100381610038161003816119908 sdot 119909119894 + 1198871003816100381610038161003816 = 1 (3)

The set of vectors is said to be optimally separated bya hyperplane if it is separated without error The distancebetween the closest vector and the hyperplane is maximalHence a separating hyperplane in canonical form mustsatisfy the following constraints

119910119894sdot (119908 sdot 119909

119894+ 119887) ge 1 119894 = 1 2 119899 (4)

Suppose that the following bound holds

1199082 le 119888 (5)

The VC dimension ℎ of the set of canonical hyperplanesin 119899 dimensional space is bounded by

ℎ le min ([1198772 119888] 119889) + 1 (6)

where 119877 is the radius of a hypersphere enclosing all trainingvectors Hence the hyperplane that optimally separates thedata is the one that minimizes (4)

120601 (119908) =1

2(119908 sdot 119908) (7)

This is equivalent to minimizing an upper bound on theVC dimension In practice such a hyperplane does not existSo we need to relax the constraints of (4) by introducing slackvariables 120585 ge 0 119894 = 1 2 119899

119910119894sdot (119908 sdot 119909

119894+ 119887) ge 1 minus 120585 119894 = 1 2 119899 (8)

In this case the optimization problem becomes

120601 (119908 120585) =1

2(119908 sdot 119908) + 119862

119899

sum119894=1

120585119894 (9)



119871 (119908 119887 120572 120585 120574) =1

2(119908 sdot 119908) + 119862

119899

sum119894=1

120585119894

minus119899

sum119894=1

120572119894

1003816100381610038161003816119910119894 (119908 sdot 119909119894 + 119887) minus 1 + 1205851198941003816100381610038161003816 minus119899

sum119894=1

120574119894120585119894

(10)





max120572

[

[

119899

sum119894=1

120572119894minus1

2

119899

sum119894119895=1

120572119894120572119895119910119894119910119895(119909119894 119909119895)]

]

(11)


sum119894=1

120572119894119910119894= 0 (12)

0 le 120572119894le 119862 119894 = 1 2 119899 (13)


120572119894[119910119894(119908 sdot 119909

119894+ 119887) minus 1] = 0 119894 = 1 2 119899 (14)


119910119894sdot (119908 sdot 119909

119894+ 119887) = 1 (15)




119891 (119909) = sgn119899

sum119894=1




120601 (sdot) 119877119889 997888rarr 119877119889ℎ (17)


max120572

[

[

119899

sum119894=1

120572119894minus1

2

119899

sum119894119895=1

120572119894120572119895119910119894119910119895119870(119909119894 119909119895)]

]

(18)

where

119870(119909119894 119909119895) = 120601 (119909

119894) sdot 120601 (119909

119895) (19)



119870(119909119894 119909119895) = (119909

119894sdot 119909119895+ 1)119902

119902 gt 0 (20)


119870(119909119894 119909119895) = exp

minus

10038161003816100381610038161003816119909119894 minus 119909119895100381610038161003816100381610038162

1205902

(21)


119870(119909119894 119909119895) = tanh (120573 sdot (119909

119894sdot 119909119895) + 119888) (22)


119891 (119909) = sgn119899

sum119894=1

120572lowast119894119910119894119870(119909119894 119909) + 119887lowast (23)


119894

120572119894[

[

119910119894

119899

sum119895=1

120572119895119910119895119870(119909119895 119909119894) + 119887 minus 1]

]

= 0 (24)






Ontology



(25)




120583FS (119911) 119911 997888rarr [0 1] (26)





FC-N r1 r2 r3 r4 rn

Fc1P3Fc1P1Fc1P2

Fc1Pn

Fc2P3Fc2P1Fc2P2

Fc2Pn

Fc3P3Fc3P1Fc3P2

Fc3Pn

FcnP3FcnP1FcnP2

FcnPn

F1s1)

F1s3)F1s2)

F1sk)

F1sn)

F2s1)F2s2)

F2s3)F2sk)

F2sn)

F3s1)F3s2)

F3s3)F3sk)

F3sn)

Fns1)

Fns2)

Fns3)

Fnsk

Fnsn)

Main category layer



Fuzzy sets layer


FC-1 FC-2 FC-3




















Object

Nonobject








Datacapturing

Dataacquisition

Datatransmitting



Featureextraction




Ontologyfeature



categorization






















6 Conclusion





BP trainingtime (s)

SVMtrainingtime (s)

Proposedsystem

training time

BP testingtime (s)

SVM testingtime (s)

Proposedsystem

testing time

BPaccuracy

SVMaccuracy



0

20

40

60

80

100


Ann(BP)Classic SVM

Proposed scheme





Acknowledgment


References

































Volume 2014




Journal of











Journal of


Function Spaces






Algebra











119871 (119908 119887 120572 120585 120574) =1

2(119908 sdot 119908) + 119862

119899

sum119894=1

120585119894

minus119899

sum119894=1

120572119894

1003816100381610038161003816119910119894 (119908 sdot 119909119894 + 119887) minus 1 + 1205851198941003816100381610038161003816 minus119899

sum119894=1

120574119894120585119894

(10)





max120572

[

[

119899

sum119894=1

120572119894minus1

2

119899

sum119894119895=1

120572119894120572119895119910119894119910119895(119909119894 119909119895)]

]

(11)


sum119894=1

120572119894119910119894= 0 (12)

0 le 120572119894le 119862 119894 = 1 2 119899 (13)


120572119894[119910119894(119908 sdot 119909

119894+ 119887) minus 1] = 0 119894 = 1 2 119899 (14)


119910119894sdot (119908 sdot 119909

119894+ 119887) = 1 (15)




119891 (119909) = sgn119899

sum119894=1




120601 (sdot) 119877119889 997888rarr 119877119889ℎ (17)


max120572

[

[

119899

sum119894=1

120572119894minus1

2

119899

sum119894119895=1

120572119894120572119895119910119894119910119895119870(119909119894 119909119895)]

]

(18)

where

119870(119909119894 119909119895) = 120601 (119909

119894) sdot 120601 (119909

119895) (19)



119870(119909119894 119909119895) = (119909

119894sdot 119909119895+ 1)119902

119902 gt 0 (20)


119870(119909119894 119909119895) = exp

minus

10038161003816100381610038161003816119909119894 minus 119909119895100381610038161003816100381610038162

1205902

(21)


119870(119909119894 119909119895) = tanh (120573 sdot (119909

119894sdot 119909119895) + 119888) (22)


119891 (119909) = sgn119899

sum119894=1

120572lowast119894119910119894119870(119909119894 119909) + 119887lowast (23)


119894

120572119894[

[

119910119894

119899

sum119895=1

120572119895119910119895119870(119909119895 119909119894) + 119887 minus 1]

]

= 0 (24)






Ontology



(25)




120583FS (119911) 119911 997888rarr [0 1] (26)





FC-N r1 r2 r3 r4 rn

Fc1P3Fc1P1Fc1P2

Fc1Pn

Fc2P3Fc2P1Fc2P2

Fc2Pn

Fc3P3Fc3P1Fc3P2

Fc3Pn

FcnP3FcnP1FcnP2

FcnPn

F1s1)

F1s3)F1s2)

F1sk)

F1sn)

F2s1)F2s2)

F2s3)F2sk)

F2sn)

F3s1)F3s2)

F3s3)F3sk)

F3sn)

Fns1)

Fns2)

Fns3)

Fnsk

Fnsn)

Main category layer



Fuzzy sets layer


FC-1 FC-2 FC-3




















Object

Nonobject








Datacapturing

Dataacquisition

Datatransmitting



Featureextraction




Ontologyfeature



categorization






















6 Conclusion





BP trainingtime (s)

SVMtrainingtime (s)

Proposedsystem

training time

BP testingtime (s)

SVM testingtime (s)

Proposedsystem

testing time

BPaccuracy

SVMaccuracy



0

20

40

60

80

100


Ann(BP)Classic SVM

Proposed scheme





Acknowledgment


References

































Volume 2014




Journal of











Journal of


Function Spaces






Algebra












Ontology



(25)




120583FS (119911) 119911 997888rarr [0 1] (26)





FC-N r1 r2 r3 r4 rn

Fc1P3Fc1P1Fc1P2

Fc1Pn

Fc2P3Fc2P1Fc2P2

Fc2Pn

Fc3P3Fc3P1Fc3P2

Fc3Pn

FcnP3FcnP1FcnP2

FcnPn

F1s1)

F1s3)F1s2)

F1sk)

F1sn)

F2s1)F2s2)

F2s3)F2sk)

F2sn)

F3s1)F3s2)

F3s3)F3sk)

F3sn)

Fns1)

Fns2)

Fns3)

Fnsk

Fnsn)

Main category layer



Fuzzy sets layer


FC-1 FC-2 FC-3




















Object

Nonobject








Datacapturing

Dataacquisition

Datatransmitting



Featureextraction




Ontologyfeature



categorization






















6 Conclusion





BP trainingtime (s)

SVMtrainingtime (s)

Proposedsystem

training time

BP testingtime (s)

SVM testingtime (s)

Proposedsystem

testing time

BPaccuracy

SVMaccuracy



0

20

40

60

80

100


Ann(BP)Classic SVM

Proposed scheme





Acknowledgment


References

































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










FC-N r1 r2 r3 r4 rn

Fc1P3Fc1P1Fc1P2

Fc1Pn

Fc2P3Fc2P1Fc2P2

Fc2Pn

Fc3P3Fc3P1Fc3P2

Fc3Pn

FcnP3FcnP1FcnP2

FcnPn

F1s1)

F1s3)F1s2)

F1sk)

F1sn)

F2s1)F2s2)

F2s3)F2sk)

F2sn)

F3s1)F3s2)

F3s3)F3sk)

F3sn)

Fns1)

Fns2)

Fns3)

Fnsk

Fnsn)

Main category layer



Fuzzy sets layer


FC-1 FC-2 FC-3




















Object

Nonobject








Datacapturing

Dataacquisition

Datatransmitting



Featureextraction




Ontologyfeature



categorization






















6 Conclusion





BP trainingtime (s)

SVMtrainingtime (s)

Proposedsystem

training time

BP testingtime (s)

SVM testingtime (s)

Proposedsystem

testing time

BPaccuracy

SVMaccuracy



0

20

40

60

80

100


Ann(BP)Classic SVM

Proposed scheme





Acknowledgment


References

































Volume 2014




Journal of











Journal of


Function Spaces






Algebra











Object

Nonobject








Datacapturing

Dataacquisition

Datatransmitting



Featureextraction




Ontologyfeature



categorization






















6 Conclusion





BP trainingtime (s)

SVMtrainingtime (s)

Proposedsystem

training time

BP testingtime (s)

SVM testingtime (s)

Proposedsystem

testing time

BPaccuracy

SVMaccuracy



0

20

40

60

80

100


Ann(BP)Classic SVM

Proposed scheme





Acknowledgment


References

































Volume 2014




Journal of











Journal of


Function Spaces






Algebra











Datacapturing

Dataacquisition

Datatransmitting



Featureextraction




Ontologyfeature



categorization






















6 Conclusion





BP trainingtime (s)

SVMtrainingtime (s)

Proposedsystem

training time

BP testingtime (s)

SVM testingtime (s)

Proposedsystem

testing time

BPaccuracy

SVMaccuracy



0

20

40

60

80

100


Ann(BP)Classic SVM

Proposed scheme





Acknowledgment


References

































Volume 2014




Journal of











Journal of


Function Spaces






Algebra























6 Conclusion





BP trainingtime (s)

SVMtrainingtime (s)

Proposedsystem

training time

BP testingtime (s)

SVM testingtime (s)

Proposedsystem

testing time

BPaccuracy

SVMaccuracy



0

20

40

60

80

100


Ann(BP)Classic SVM

Proposed scheme





Acknowledgment


References

































Volume 2014




Journal of











Journal of


Function Spaces






Algebra












BP trainingtime (s)

SVMtrainingtime (s)

Proposedsystem

training time

BP testingtime (s)

SVM testingtime (s)

Proposedsystem

testing time

BPaccuracy

SVMaccuracy



0

20

40

60

80

100


Ann(BP)Classic SVM

Proposed scheme





Acknowledgment


References

































Volume 2014




Journal of











Journal of


Function Spaces






Algebra































Volume 2014




Journal of











Journal of


Function Spaces






Algebra









research article an obstacle recognizing mechanism for...

Documents