Gait Recognition using SVM and LDA by idesajith


									Short Paper
                       Proc. of Int. Conf. on Advances in Computing, Control, and Telecommunication Technologies 2011

                    Gait Recognition using SVM and LDA
                                    Saeid Fazli, Hadis Askarifar, Mahmoud Joz Tavassoli
                               Electrical Eng. Dept., Eng. Faculty, Zanjan University, Zanjan, Iran

Abstract: Biometric information-based human recognition is                  invasive, unobvious, low resolution requirement. Furthermore
important due to its reliability in identity verification. Human            it is the only noticeable biometric feature for human
identification at a distance is an attractive task in visual                identification at a distance. Gait recognition algorithms are
surveillance. Gait recognition has this ability to recognize
                                                                            influenced by specific views of camera, time of testing,
individuals from a distance [1]. Gait recognition has 3 steps.
The first step is preprocessing, the second step is feature                 clothing, health status, stance and posture of person [5].
extraction and the third one is classification. In this paper, we               This paper is organized as follows: In section 2 we explain
closely focus on the three steps, At the first step, data are               dataset standardization method. In Section 3 feature extraction
standardized as described in this paper ,then LDA(Liner                     is described that is the base of temporal changes of the
Discriminant Analysis) is utilized for feature reduction and                walker’s silhouette and features fusion. In Section 4, we pay
finally, Multi Class SVM (Support Vector Machine) classifier                attention to Linear Discriminant Analysis (LDA) to represent
is used. Databases are used in our experiments. Experimental                the original gait features from a high-dimensional
results show the effectiveness of our proposed method in gait               measurement space to a low-dimensional eigenspace. Section
                                                                            5 represents the SVM classification technique as an optimal
Keywords: Human motion analysis, biometrics, gait                           discriminant method, based on the Bayesian learning theory.
recognition, Liner Discriminant Analysis (LDA), Multi class                 The proposed method is explained in section 6. Section 7
Support Vector Machine (SVM).                                               shows the experimental results in detail. Finally, conclusions
                                                                            form the last section.
                        I. INTRODUCTION
                                                                                             II. DATASET STANDARDIZATION
Visual surveillance is important in public places such as
airports, shopping malls and banks. Therefore biometric                         In our database we have different views that have different
information of every person is necessary for human                          silhouette in particular in person’s height and width. Another
identification and recognition. In the past, the field of biometric         problem is many extra pixels in the initial silhouette. The main
researches has focused on information resulted from finger                  goal is producing a dataset like the one that a person is walking
prints, shoe prints, iris images, palm print and suffers from               on a treadmill i.e. the same position of the person in the middle
some disadvantages: 1) As cameras covered the whole area,                   of each frame and the same size in the whole image sequence.
persons are usually far from the cameras. Therefore many                    The idea is to fix the head for each frame in a predefined
above mentioned features can not be captured. In fact, it can               position and resize the body to achieve a preset height. In
be easily determined that in above cases obtaining biometric                this regard, we perform a three stage preprocessing. Firstly,
information in detail is nearly unachievable due to the distance            we extract a rectangle including just the person without extra
of camera to individuals. 2) Person’s partnership: In most                  black pixels and obtain height and width of the person. In the
time person’s partnership for collecting biometric information              sequence is calculated and each frame is converted to the
is necessary. Participation of every one to collecting these                biggest height and width. Finally, we move the head of each
types of biometric information is necessary. In fact, majority              frame in a fixed point .The process is illustrated in Fig. 1.
of current fingerprint or iris databases in airports or companies
have been prepared with the cooperation of cooperators.
More over to achieve each of this biometric information,
variety of instruments are needed. 3) Using respectful
methods for individuals. Sometimes system may be out of
order and make people to do repetitive actions, thereby
slowing things down, these methods of collecting biometric
information is not supposed as a respectful way, as a whole.
The best solution for above issues is a new behavioral
biometric named gait recognition which focus on identifying
the person by the way he or she walks. As a matter of fact,
gait is affected by weight, limb length, habitual posture, bone
structure, age and health status and all these parameters are                           Fig. 1.Three steps of standardizing dataset
unique for everybody. As a result gait can be an individual
feature for person recognition. Compared with the mentioned
generational biometrics (face, fingerprints and iris) gait owns
great CONSPICUOUS benefits of being non-contact, non-
© 2011 ACEEE
DOI: 02.ACT.2011.03.42
Short Paper
                      Proc. of Int. Conf. on Advances in Computing, Control, and Telecommunication Technologies 2011

                      III. FEATURE EXTRACTION                            This demonstrates shape’s information in 1D space. Finally,
                                                                         we can convert these distance signals to normalized ones
    In our previous work ([12]), each image frame at first step
                                                                         with considering magnitude and size. In the first place, its
is converted into an associated temporal sequence of distance
                                                                         signal magnitude will be normalized through L1- norm. Then,
signals at the preprocessing stage. Temporal variations of
                                                                         equally spaced re-sampling is used to normalize its size into
the walker’s silhouette are a clear-cut point in determining
                                                                         a fixed length (360 in our experiments). By converting such a
basic motion of a walking figure. Neutralizing the proposed
                                                                         sequence of silhouette images into an associated sequence
method’s changes pertaining to color and texture of clothes,
                                                                         of 1D signal patterns, we will no longer need to cope with
only the binary silhouette has been selected. In addition,
                                                                         those likely noisy silhouette data [2].
with converting 2D silhouette changes into 1D signal we can
achieve higher computational efficiency. In our new work we
                                                                                       IV. DIMENSION REDUCTION (LDA)
apply this method on “OR” figure that achieve from “OR”
operator performance on images sequence. These stages are                     One of the most powerful techniques in dimensionality
showed in Fig.2. In our database we have moving silhouette               reduction is LDA which is an acronym for Linear Discriminant
that has been tracked second stage, the biggest height and               Analysis [7]. Maximizing the linear separability between data
width in a frame from a walking figure, then using a suitable            points associated with different classes is the main goal in
algorithm which follows the border, outer contour can be                 LDA. In fact, in order to maximize the linear class separability
perceived. Therefore, we must compute centroid shape (x c,               in the low-dimensional illustration of data the method finds a
yc). After converting outer contour to a distance signal, we             linear mapping M [8]. SW and Sb which are referring to the
travel it clockwise. Every element of the distance signal S=             within-class scatter and the between-class scatter are two
{d1, d2, di,…,dNb} is distance between point of outer contour            criteria used to formulate linear class separability in LDA as
and the shape’s centroid.                                                follows:

 Fig.2. “ OR” Silhouette representation: (a) illustration of boundary extraction and counter clockwise unwrapping and (b) the normal-
                  ized distance signal consisting of all distance between the centroid and the pixels on the boundary
formulate linear class separability in LDA as follows:

                                                                         This maximization can be performed by solving the generalized
                                                                         eigen problem

                                                                         For the d largest eigenvalues (under the requirement that d <
 Where                                                                   |C|). The eigenvectors v forms the columns of the linear
 Pc: the class prior of class label c.                                   transformation matrix T. The low-dimensional data
                 : The covariance matrix of the zero mean data           representation Y of the data points in X can be computed by
point’s xi assigned to class c C (where C is the set of possible         mapping them onto the linear basis T, i.e., Y =     )T. LDA
classes).                                                                has been successfully applied in a large number of
           : The covariance matrix of the cluster means.                 classiûcation tasks. Successful applications include speech
               : Covariance matrix of the zero means data X.             recognition [9], mammography [10], and document
LDA optimizes the ratio between the within-class scatter Sw              classiûcation [11].
and the between-class scatter Sb in the low-dimensional
representation of the data, by ûnding a linear mapping that
maximizes the so-called
Fisher criterion
© 2011 ACEEE
DOI: 02.ACT.2011.03.42
Short Paper
                       Proc. of Int. Conf. on Advances in Computing, Control, and Telecommunication Technologies 2011

                        V. CLASSIFICATION                                   Must be solved in order to obtain the vector       and the
    One of the successful techniques for classification is SVM              Scalar . Successively, the classification function
[13, 14, and 15], so we use SVM classifiers here. SVM is an
optimal discriminant method, based on the Bayesian learning                 is used to discriminate between the two sets of elements [17].
theory. For the cases where it is difficult to estimate the density             SVMs were originally designed for binary classification.
model in high dimensional spaces, the discriminant approach                 Some methods have been proposed where typically construct
is preferable to the generative approach. SVM performs an                   multi-class classifier by combining several binary classifiers
implicit mapping of data into a higher dimensional feature                  like one-against-all, it construct k SVM models where k is the
space, and then finds a linear separating hyper plane with                  number of classes. The ith SVM is trained with all of examples
the maximal margin to separate data.                                        in the ith class with positive labels, and all other examples
     In this higher dimensional space [16].In the context of                with negative labels. Another major method is called the one
pattern classifying, the main purpose is to find the optimal                against one method. This method constructs k (k-1)/2
separating hyper-plane, that is, the hyper plane that separates             classifiers where each one is trained on data from two classes.
the positive and negative examples with maximal margin. In
fact, a good choice is a hyper-plane that leaves maximum                                        VI. PROPOSED METHOD
margin between two classes, where the margin is defined as
sum of the distances of the hyper-plane from closest point of                   Fig.3 shows a block diagram of the proposed algorithm.
the two classes. The goal of SVMs is to produce a model that                The algorithm consists of preprocessing, data
predicts target values L for new data instances. More                       standardization, feature extraction and fusion and finally SVM
specifically, given L samples training data set of L samples:               classification. We propose fusion of two features for
                                                                            improving the results. At first we apply “OR” operator on
                                                                            every frame sequence for each person as explained in section
               and a function                            the                2. Unwrapping the outer contour of the resulted shape is
optimization problem:                                                       done in the second stage. In addition, The LDA method, as
                                                                            described by Belhumeur et al. [7], uses both PCA and LDA to
                                                                            produce a subspace projection matrix, minimizing within-class
                                                                            variation and maximizing between-class variation. The LDA
                                                                            method is performed in the second step. Finally, SVM as a
                                                                            successful technique for classification is applied at the last

                               Fig.3. The block diagram of the proposed multimodal gait recognition system
                                                                            and an evaluation framework can be found in this paper. The
                         VII. E XPERIMENTAL RESULT                          format of the video filename in Dataset B is ‘xxx-mm-nn-ttt.avi’,
    We conducted some experiments on the CASIA Gait
                                                                                  xxx: subject id, from 001 to 124.
Database (Dataset B) [3]. In the CASIA Gait Database there
                                                                                  mm: walking status can be ‘nm’ (normal), ‘cl’ (in a
are three datasets: Dataset A, Dataset B (multiview dataset)
                                                                                      coat) or ‘bg’ (with a bag).
and Dataset C (infrared dataset). Dataset B is a large multiview
                                                                                  nn: sequence number.
gait database, which is created in January 2005. There are 124
                                                                                  ttt: view angle, can be ‘000’, ‘018’, ... , ‘180’.
subjects, and the gait data was captured from 11 views. Three
                                                                            There are 13,640 (124×10×1) video sequences in our data-
variations, namely view angle, clothing and carrying
                                                                            base, with 2–3 gait cycles our each sequence. The frame size
condition changes, are separately considered. Besides the
                                                                            is 320-by-240 pixel, and the frame rate is 25 fps. We applied
video files, has still provided human silhouettes extracted
                                                                            %15 of persons in this database to have the same scenario as
from video files. The detailed information about Dataset B
© 2011 ACEEE
DOI: 02.ACT.2011.03.42
Short Paper
                       Proc. of Int. Conf. on Advances in Computing, Control, and Telecommunication Technologies 2011

of the work of L. Wang et al. The achieved results using the                                        REFERENCES
proposed method has better CCR compared the one of [2].
                                                                         [1] M. Pushpa Rani , G. Arumugam “An Efficient Gait Recognitio
Based on the experimental results, fourth, fifth and sixth views         System Human Identification Using Modified ICA”,International
(equal to 54, 72 and 90 degrees) have better CCR results and             jornal of computer science& information Technology (IJCSIT)
the best one is 90 degree lateral view. The results are shown            ,Vol.2,No.1, February2010.
in the tables 1. Table 1 reveals that the obtained results are           [2] S. Niyogi and E. Adelson, “Analyzing and recognizing walking
better than [2, 4]. The database used in this work (database             ûgures in XYT,” in Proc. IEEE Comput. Vis. Pattern Recog. Seattle,
B) is a very large database, rarely used by researchers and              WA, 1994, pp. 469–474.
the results show the effectiveness of the proposed method.               [3] J. E. Boyd, “Synchronization of oscillations for machine
                                                                         perception of gaits,” Comput. Vis. Image Understanding, vol. 96,
                                                                         no. 1, pp. 35–59,2004.
                       VIII. CONCLUSIONS
                                                                         [4] D.Gavrila, “The visual analysis of humanmovement: Asurvey,”
    In this paper, we use OR contour as features and LDA                 Comput. Vis. Image Understanding, vol. 73, pp. 82–98, 1999.
and SVM for efficient gait recognition. The proposed                     [5] X.Li , S j. Maybank and et al.”Gait Components and Their
algorithm results in better recognition accuracy. Based on               Application to Gender Recognition “, IEEE TRANSACTIONS
                                                                         ON SYSTEMS, MAN, AND CYBERNETICS—PART C:
the experimental results, fourth, fifth and sixth views (equal
                                                                         APPLICATIONS AND REVIEWS, VOL.38,NO.2,MARCH2008.
to 54, 72 and 90 degrees) have better CCR results and the                [6] L. Wang, T. Tan, H. Ning, and W. Hu, “Silhouette Analysis-
best one is obtained for the 90 degree lateral view. The results         Based Gait recognition for Human Identification”, IEEE
are shown in table 1. The table reveals that the proposed                Transactions on Pattern Analysis and Machine Intelligence, Vol.
algorithm works better than [6, 12] i.e. around 20% increase             25, pp. 1505-1518, 2003.
in CCR. This is because of a novel standardizing method and              [7] P. Belhumeur, J. Hesponha, and D. Kriegman. Eigenfaces vs
applying contours on the OR result of the whole image                    Fisherfaces: Recognition using class speciûc linear
sequence instead on one image and using LDA and SVM for                  projection. IEEE Trans. Pattern Analysis and Machine Intelligence
feature reduction and classification.                                    , 18, 1996.
                                                                         [8] R.A. Fisher. The use of multiple measurements in taxonomic
             TABLE I. COMPARED RESULTS   FOR LATERAL VIEW                problems. Annals of Eugenics, 7:179–188, 1936.
                                                                         [9] R. Haeb-Umbach and H. Ney. Linear discriminant analysis for
                                                                         improved large vocabulary continuous speech recognition. In IEEE
                                                                         International Conference on Acoustics, Speech, and Signal
                                                                         Processing, 1992, volume 1, pages 13–16, 1992.
                                                                         [10] H.-P. Chan. Computer-aided classiûcation of mammographic
                                                                         masses and normal tissue: linear discriminant analysis in texture
                                                                         feature space. Phys. Med. Biol., 40:857–876, 1995.
                                                                         [11] K. Torkkola. Linear discriminant analysis in document
                                                                         classiûcation. In IEEE ICDM-2001 Workshop on Text
                                                                         Mining, pages 800–806, 2001.
                                                                         [12] Saeid F, Hadis A, Maryam S, “Multi-View Neural Network
                                                                         Based Gait Recognition “,World Academy of Science, Engineering
                                                                         and Technology 67 2010.
                                                                         [13] B. Moghaddam, M. Yang, Learning gender with support faces,
                                                                         IEEE Trans.Pattern Anal. Mach. Intell. 24 (5) (2002) 707–711.
                                                                         [14] L. Lee, W.E.L. Grimson, Gait analysis for recognition and
                                                                         classification, in: IEEE International Conference on Automatic Face
                                                                         & Gesture Recognition (FG),2002, pp. 155–162.
                                                                         [15] L. Walawalkar, M. Yeasin, A. Narasimhamurthy, R. Sharma,
                                                                         Support vectorlearning for gender classification using audio and
                                                                         visual cues, Int. J. Pattern Recognition Artif. Intell. 17 (3) (2003)
                                                                         [16] G. Shan,S. Gong and P W. McOwan ,” Fusing gait and face
                                                                         cues for human gender recognition”, Neurocomputing (2008).
                                                                         [17] C. Hsu, C. Lin (2002) “A Comparison of Methods for
                                                                         Multiclass Support Vector Machines”, IEEE Trans. NEURAL
                                                                         NETWORKS, Vol. 13, NO.2, MARCH 2002.

© 2011 ACEEE
DOI: 02.ACT.2011.03.42

To top