gait recognition using mda, lda, bpnn and svm

Int. Journal of Electrical & Electronics Engg. Vol. 2, Spl. Issue 1 (2015) e-ISSN: 1694-2310 | p-ISSN: 1694-2426

NITTTR, Chandigarh EDIT -2015 92

Gait Recognition using MDA, LDA,BPNN and SVM

1RishamPuri, 2Jatinder Kumar, 3Rakesh Kumar1Student M.Tech (CSE) SSCET, Manawala (Amritsar), 2,3A.P. (CSE) SSCET, Manawala (Amritsar),

[email protected],[email protected],[email protected]

Abstract: - Recognition of any individual is a task to identifythe human beings. Human identification using Gait is methodto identify an individual by the way he walk or manner ofmoving on foot of humans. Gait recognition is a type ofbiometric recognition and related to the behavioralcharacteristics of biometric recognition. Gait offers ability ofdistance recognition or at low resolution. In this paper it willpresent the review of gait recognition system where differentapproaches and classification categories of Gait recognitionlike model free and model based approach, MDA, BPNN,LDA, and SVM.

Keywords: -Gait Recognition, Back Propagation NeuralNetwork (BPNN), SVM, MDA, LDA and identification.

I. GAIT RECOGNITIONThe identification through biometric is a better waybecause it associate with individual not with informationpassing from one place to another. The biometric is a fieldof technology that uses automated methods for identifyingand verifying a human. In real time applications like inbanks; airports; authentications and verifications arealways required. In such type of applications biometricidentification methods are used [1].The biometric has two main characteristics:A. Physiological:These are biometrics which is derived from a directmeasurement of a part of a human body. Then mostprominent and successful of these types of measures thatare Face, fingerprints, iris, palm print, DNA etc. These arerelated to body.B. Behavioural:Voice and Gait are related to behaviour of the person.Extract characteristics based on an action performed by anindividual; they are an indirect measure of thecharacteristic of the human form. The main feature of abehavioural biometric is the use of time as a metric. Thenestablished measures include keystroke-scan and speechpatterns. Biometric identification should be an automatedprocess. Therefore manual feature extraction would beboth undesirable and time consuming; due to the largeamount of data that must be acquired and processed inorder to produce a biometric signature. And inability toautomatically extract the desired characteristics whichwould render the process infeasible on realistic size datasets in a real-world application.C. Gait Analysis:Gait analysis is the systematic study of human locomotion;augmented by instrumentation for measuring bodymovements; body mechanics and the activity of themuscles [2]. Gait based recognition is more suitable invideo surveillance applications because of followingadvantages:

1. Recognition using gait do not need any user cooperation.2. The gait of an individual can be captured at a distance.3. Gait recognition does not require images of veryHigh quality and provide good results in low resolution.

D. Approaches for Gait Recognition:Some basic methods and approaches for gait recognition[3]:D.1. Moving Video based gait recognition:In this approach, gait is captured using a video-camerafrom a distance. Image and video processing techniquesare employed to extract gait features for the purpose ofrecognition. For example stride, cadence, static bodyparameters extra.D.2. Floor Sensor based gait recognition:In this approach, a set of sensors or force plates areinstalled on the floor and such sensors enable to measuregait related features, when a person walks on them, e.g.maximum time value of heel strike and maximumamplitude value of the heel strike extra.D.3. Wearable Sensor based gait recognition:In this approach, gait is collected using body worn motionrecording (MR) Sensors on human body. The MR sensorscan be worn at different locations on the human body. Theacceleration of gait, which is recorded by the MR sensor, isutilized for authentication [4, 5].

E. Steps of Gait Recognition SystemE.1. the Background Subtraction:In this approach moving objects from background in thescene are identified first. Then some of the backgroundsubtraction techniques are applied on it .A commonapproach is to perform background subtraction; whichidentifies moving objects from the portion of video framethat differs from the background model. The backgroundsubtraction generates binary images containing black andwhite (moving pixels) also known as binary silhouettes.The background subtraction is a class of techniques forsegmenting out objects of interest in a scene forapplications such as surveillance. Therefore there are manychallenges in developing a good background subtractionalgorithm. 1st it must be robust against changes inillumination task. 2nd it should avoid detecting non-stationary background objects such as moving leaves; rain;snow and shadows cast by moving objects. And finally; itsinternal background model should react quickly to changesin background such as starting and stopping of vehicles.E.2. Pre-processing:Pre-processing is done on video frames to reduce presenceof noise then some filters are applied which in turns blurthe frames of image, which helps in shadow removal, afterpre-processing motion detection is performed. Background


93 NITTTR, Chandigarh EDIT-2015

subtraction technique uses the difference of current imageand background to detect the motion. It delineates theforeground from background. Background subtractiongenerate binary image containing black (background) andwhite (moving pixel) then post processing is applied toobtain normalized silhouette images with less noise. Theyused morphological operators such as dilation and erosionto fill small holes inside silhouette and to filter small noiseon the background. To reduce computational cost theyproposed new silhouette representation method which onlyuses some of pixel on the contour [6].E.3. Feature Extraction:Feature extraction is a special form of dimensionalityreduction. And when the input data is too large to beprocessed and it is suspected to be notoriously redundant(e.g. the same measurement in both feet) then the inputdata will be transformed into a reduced representation setof features (also named features vector). Then transformingthe input data into the set of features is called featureextraction.E.4. Recognition:This is the final step of human identification using gait. Inthis step input videos are compared with sequences storedin database. Different types of classifiers are used for therecognition. Such as: MDA (Multi-linear Discriminantanalysis) and LDA (Linear Discriminant Analysis).Theyuse MDA approach to optimize the separability of gaitfeatures [7].

F. Gait Recognition SystemSystem will identify unauthorized individual and comparehis gait with stored sequences and recognize. Thebackground subtraction is the common approach of gaitrecognition.Using background subtraction, pre-processing is done toreduce noise. The background subtraction techniques arealso classified into two types: non- recursive methods andrecursive methods. Non recursive techniques use slidingwindow approach for background subtraction. Therecursive methods use single Gaussian method andGaussian mixture model. The Gait recognition methodcontains two parts1. Training part2. Testing partGait analysis laboratory has several cameras (video orinfrared) placed around treadmill. Then person has markerslocated at various points of body (e.g. spines of the pelvis,ankle malleolus). When person walks down the treadmilland the computer calculates the trajectory of each markerin three dimensions. And model is applied to calculate themovement of bones.

Applications of gait:Gait recognition technology is not limited to securityapplications researchers also envision medicalapplications. For example, recognizing changes in walkingpatterns early on can help to identify conditions such asParkinson’s disease and multiple sclerosis in their earlieststages.Medical diagnostics: In computerized gait analysis andpatient walks or run with sensors in his foot. The sensorsends some points of info- about foot pressure and timingand range of motion to computer and creates diagram.Doctor can review them and came up with treatment plan.

Biometric identification and forensics: Gait Pal and PalEntropy Minor variations in gait style can be used asa biometric identifier to identify individual people [8].

II. BPNNIn this paper, we use one classical type of neural networks–BPNN. BPNN usually has input and output layers withsome hidden layers. Actually BPNN can be likened to aflexible mathematical function which has manyconfigurable internal parameters to find the results. Inorder to accurately represent the complicated relationshipsamong gait variables and these internal parameters need tobe adjusted through training process. In training processgait features and corresponding labels are input to thenetwork, which iteratively self-adjusts to accuratelyclassify as many gait features as possible. Training iscomplete when some criterion is satisfied (e.g., interactiontimes reach a preset value or prediction error falls below apreset threshold). Once the neural network is trained wecan use it to predict the gait features of sequences of gaittesting. It is to be noted that the trained neural networksimply performs function evaluation using the internalparameters established during training process to producean output [9].

III. SUPPORT VECTOR MACHINEThe theory of SVM is based on the idea of structural riskminimization. In many applications SVM has beenintroduced as a powerful tool for solving classificationproblems. There are many researchers have used SVM ongait recognition. It is to be noted that SVM isfundamentally a classifier of two-tier. SVM first maps thetraining samples into a high dimension space (typicallymuch higher than the original data space) and then finds aseparating hyper plane that maximizes the margin betweentwo classes. Maximizing the margin is a quadraticprogramming (QP) problem and can be solved from itsdual problem by introducing Lagrangian multipliers oftechnique. Without any knowledge of the mapping theSVM can find the optimal hyper plane by using the dotproduct functions in original space that are called kernelsof image. There are several kernels proposed byresearchers. Here we use radial basis function (RBF). Oncethe optimal hyper plane is established we can directly use adecision function to classify testing samples. For solvingmulti-class problems and various methods have beenproposed for combining multiple two classes SVMs inorder to build a multi-class classifier such as one-against-one and one-against-rest method. In this paper we use theone against-one method in which k (k*1) =2 classifiers areconstructed and each one trains samples. In classificationwe use a voting strategy: each two-class SVM isconsidered as a voter (i.e. k (k * 1) =2 voters in all) andthen each testing sample is classified to the class withmaximum number of votes [10].

IV. MULTILINEAR DISCRIMINANT ANALYSISThe linear Discriminant analysis (LDA) is a classicalalgorithm that has been successfully applied and extendedto various biometric signal recognition problems. Therecent advancement in multi-linear algebra led to a numberof multi-linear extensions of the LDA, multi-linear


NITTTR, Chandigarh EDIT -2015 94

Discriminant analysis (MLDA), being proposed for therecognition of biometric signals using their naturaltonsorial representation. In general, MLDA seeks a multi-linear projection that maps the input data from one space toanother (lower dimensional, more discriminative) space.

V. LINEAR DISCRIMINANT ANALYSISLinear Discriminant Analysis (LDA) is a well-knownscheme for feature extraction and reduction in dimension.It has been used widely in many applications involvinghighly dimensional data, such as face recognition andimage recognition. Linear Discriminant Analysis easilyhandles the case where the within-class frequencies areunequal and their performances have been examined onrandomly test data that are generated. This methodmaximizes the ratio of between-class variance to thewithin-class variance in any particular data set therebyguaranteeing maximal separability. We decided toimplement an algorithm for LDA in hopes of providingbetter classification compared to Principal ComponentsAnalysis. Linear Discriminant Analysis (LDA) is atechniques used for data classification and dimensionalityreduction in this section we give a brief overview ofclassical LDA.

= ∑ ∑ ( − )∏ ( − )) … (1.1)And

= ∑ ( -m) ( -m) … (1.2)where

= ∑ ∏ is the mean of the ith class… (1.3)and

m = ∑ ∑ ∏ x is the global mean … (1.4)

In Discriminant analysis, two scatter matrices, calledwithin-class (Sw) and between-class (Sb) matrices aredefined to quantify the quality. Linear DiscriminantAnalysis is a well-known scheme for feature extraction anddimension reduction. It has been used widely in manyapplications such as face recognition, image retrieval,microarray data classification, etc. The LDA method isemploys to perform training and projecting on original gaitfeature. They decrease dimensionality of high dimensionalfeature with PCA, and perform optimal classification onlow dimensional space with the LDA algorithm. Theobjective of LDA is to perform dimensionality reductionwhile preserving as much of the class discriminatoryinformation as possible [10].

VI. RESULTSIn the following figures, result of proposed algorithm ishighlighted.

Figure 1: Correct Classification Rate

Figure 2: Comparison of CCR between previous and proposed work

Figure 3: Matching Results

walk. Therefore Several Parameters has been proposed forGait Recognition previously but there have been alwaysneed for better parameters to improve recognition. Theexisting Gait Recognition Technique in doesn't considerthe distance between hands as parameters as we areconsidering this. Thus propose an Enhanced GaitRecognition Technique which is based on model basedapproach. The existing Correct Classification Rate is poor.They gave their better CCR results using SVM technique.They are less accurate and needs enhancement by BPNNtechnique. Our objective is to obtained better result usingBPNN +SVM + MDA and LDA technique.

ACKNOWLEDGMENTThanks to my Guide and family member who alwayssupport and guide me during my dissertation. Specialthanks to my father who always support my innovativeideas.

REFERENCES

VII. CONCLUSIONGait recognition aims to identify people by the way they

[1] Lili Liu, Yilong Yin, Wei Qin & Ying Li, 2011. “Gait RecognitionBased on Outermost Contour”, International Journal of Computational Intelligence Systems International Journal of Computational IntelligenceSystems, Vol. 4, No. 5.[2] Shalini Agarwal, Shaili Mishra, 2012. ” A study of multiple human tracking for visual surveillance”, Department of CS, Banasthali University, Rajasthan, International Journal of Advances in Engineering & Technology.[3] Alese, B. K., Mogaji, S. A., Adewale, O. S. and Daramola, O., 2012. ” Design and Implementation of Gait Recognition System”, Department of Computer Science, Federal University of Technology, Akure, Nigeria, International Journal of Engineering and Technology Volume 2 No. 7.[4] M.Pushparani & D.Sasikala, 2012. “ A Survey of Gait Recognition Approaches Using PCA & ICA”, Mother Teresa Women’s University, Kodaikanal, India, Global Journal of Computer Science and Technology Volume 12 Issue 10 Version 1.0.[5] Liang Wang, Tieniu Tan, Huazhong Ning, and Weiming Hu, 2003. “Silhouette Analysis-Based Gait Recognition for Human Identification”, IEEE transactions on pattern analysis and machine intelligence, VOL. 25, NO. 12.[6] Sanjeev Sharma, Ritu Tiwari, Anupam shukla and Vikas Singh, 2011. “Identification of People Using Gait Biometrics” International Journal of Machine Learning and Computing, Vol. 1, No. 4.


95 NITTTR, Chandigarh EDIT-2015

[7] M.K. Bhuyan and Aragala Jagan, 2010. “Person Identification usesGait by Combined Features of Width and Shape of the Binary Silhouette”,International Journal of Computer and Information Engineering.

[8] Qiong Cheng, Bo Fu, and Hui Chen, 2009. “Gait Recognition Basedon PCA and LDA“, School of Electrical & Electronic Engineering, HubeiUniversity of Technology Proceedings of the Second SymposiumInternational Computer Science and Computational Technology(ISCSCT’09) Huangshan, P. R. China, pp. 26-28.

[9] A. Hayder, J. Dargham, A. Chekima, and G. M. Ervin, 2011. “PersonIdentification Using Gait”, International Journal of Computer andElectrical Engineering, Vol. 3, No. 4.

[10] Amit Kale, Aravind Sundaresan, A. N. Rajagopalan, Naresh P.Cuntoor, Amit K. Roy-Chowdhury, Volker Krüger, and Rama Chellappa,2004. “Identification of Humans Using Gait” IEEE transactions on imageprocessing, VOL. 13, NO. 9.

gait recognition using mda, lda, bpnn and svm

Documents