Performance Evaluation of SVM based Abnormal Gait Analysis with Normalization by ijcsis


More Info
									                                                               (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                         Vol. 8, No. 5, 2010

               Performance Evaluation of SVM based
             Abnormal Gait Analysis with Normalization
                     M. Pushpa Rani1                                                               G.Arumugam2
           Associate Professor in Computer Science,                            2
                                                                                   Professor & Head, Dept. of Computer Science,
             Mother Teresa Women’s University,                                             Madurai Kamaraj University,
               Kodaikanal, Tamil Nadu, India                                               Madurai, Tamil Nadu, India

Abstract—Support Vector Machine classifiers are powerful tools              gait data to describe the human motion; walking speed, joint
that are specifically designed to solve large-scale classification          angles, forces, and moments etc., Data like joint kinetics, joint
problems. In 1990s, Vapnik along with a group of other                      moments and joint powers have also been used for gait
mathematicians and scientists developed a new statistical
                                                                            recognition. Also this technology is very useful for checking
approach that is more efficient particularly in dealing with large
                                                                            the walking pattern of children as Children under the age of 13
classification problems which they called as Support Vector
Machines (SVM). An SVM method is being broadly used in gait                 have more chances to have different style of walking. To trace
analysis because of its remarkable learning ability. In this paper,         out any abnormality in children’s walk, our proposed method
a two stage SVM algorithm is proposed for children abnormal                 takes a key role; and with this one could diagnose any existing
gait analysis. The algorithm uses T-Test based preprocessing                fault features of walking in early ages of childhood itself. This
methods for feature selection, normalization and combines SVM               will surely of great help for earlier treatment of gait
for Classification. Only samples that have weak relationships               abnormality in children.
with all the clusters are involved in SVM. Experimental results                An SVM method has been broadly used in gait analysis
reveal that this algorithm based on T-Test-SVM combination
                                                                            because of its remarkable learning ability, accuracy and
achieves a remarkable recognition performance for children
                                                                            efficiency. In this paper the SVM technique is performed by
abnormal gait analysis with reduced Computational cost.
                                                                            having a training set and test samples. The training set is
Keywords Abnormal Gait Analysis, Support Vector Machine                     categorized into different sets of conditions, which can be
(SVM), Gait Data Classification, T-Test                                     grouped into two classes i.e. normality and abnormality. For
                                                                            this, the collection of data is very important and several
                                                                            observations are needed. The information regarding leg length,
                       I.    INTRODUCTION
                                                                            height, cadence, stride length and age are some of the features
GAIT analysis is very significant for early diagnosis of gait               which are of great help in this gait analysis. The SVM
diseases and treatment assessment. Doctors in earlier days                  constructs a hyper plane or a set of hyper planes in a high or
used to diagnose gait diseases manually with the help of                    infinite dimensional space, which can be used for
certain graphs generated by the gait analysis system, with                  classification or clustering. In simple words, given a set of
which only vague clues which may or may not reflect the                     training examples, each with a label of belonging to one of
reality were obtained. In most cases, doctors had different                 these categories, an SVM training algorithm builds a model
views of opinion by interpreting these curves. As machine                   that predicts in which category a new example falls. The
based learning technology has developed, it has gained much                 proposed method uses t-test-SVM for classification.
interest in gait analysis, which is of great support to doctors             Interestingly, a good separation is achieved by the hyper plane
for more reliable and accurate diagnosis of a disease. A Gait               that has the largest distance to the nearest training data points
analysis is a systematic study of human motion. i.e., walking,              of any class. The t-test method is used to normalize data prior
running, skipping and the like, which mainly concentrates on                to classification. In its simplest form t-test provides a
the physical activities. Gait analysis is very much useful to               statistical analysis of means of several groups and therefore
check out the human conditions; whether normal or abnormal                  can generalize the Student's two-sample t-test to more than
using the eye and brain of the observers, augmented by                      two groups.
instrumentation for measuring body movements, body                             The following section of this paper is projected as follows:
mechanics and the activity of the muscles. In other words, gait             Section 2 discusses some of the related works done earlier in
analysis is used to assess, to plan and to treat the individuals            gait based classification. The proposed t-test SVM gait
with conditions affecting their ability to walk. There exist a lot          classification method is described in Section 3. Section 4
of methods to check out these normality and abnormality on                  illustrates the performance Analysis and Section 5 concludes
human gait, but SVM is found to be more suitable in terms of                the paper with directions for future work.
its efficiency [2]. Motion analysis provides large volume of

                                                                                                        ISSN 1947-5500
                                                             (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                       Vol. 8, No. 5, 2010
                    II.   RELATED WORK                                       J.H. Yoo et al., [18] had described an automated gait
The study of human gait has created much interest in many                 recognition system using back propagation neural network
application areas including biomechanics, clinical analysis,              algorithm. Gait is the most common human motion, and each
computer animation and biometrics. As a result many                       person appears to have his or her own characteristic gait
researches emerged in recent years and of which, a few studies            pattern. To identify the human gait, a total of 27 parameters
related to Gait classification are noted herewith.                        are considered as gait features. By calculating a class
   C. Bauckhage et al., [13] well thought-out about automatic             separability of the given feature, only 10 important features for
gait analysis as a means to deduce if an observed walking                 classifying the gait are selected from these feature sets. Then,
pattern appears to be normal or not. When compared to most                the enhanced back-propagation neural network algorithm is
contributions to visual gait analysis, the problem dealt with the         applied to the SOTON database, and recognition rate of 90%
paper requires a representation that abstracts from individual            for 30 subjects is accomplished. The results achieved give
gait characteristics but allows for the classification of gait            promising performance and higher recognition rates than those
across individuals. Addressing this requirement, the author               of an earlier gait recognition approach.
presented a homeomorphism between 2D lattices and shapes                     Ju Han et al., [19] proposed a new spatio-temporal gait
that enables a robust vector space embedding of silhouettes.              representation, called the Gait Energy Image (GEI), for
Sampling apt lattice points allows to roughly track the                   individual recognition by gait. Different from other gait
movement of limbs without requiring any limb recognition                  representations which consider gait as a sequence of templates
strategy. Combining shape representations obtained from                   (poses), GEI represents human motion sequence in a single
several frames into lager feature vectors provides temporal               image while preserving temporal information. To overcome
context for the classification task. Experimental results expose          the limitation of training templates, a simple model is
a complete knowledge that gait classification using support               proposed for simulating distortion in synthetic templates and a
vector machines yields excellent accuracy. Temporal filtering             statistical gait feature fusion approach for human recognition
of the results of classification in further improvements of the           by gait. Experimental results show that a) GEI is an effective
reliability of the presented framework, because it lessens the            and efficient gait representation and b) the proposed
effect of sporadic misclassifications.                                    recognition      approach     achieves    highly    competitive
   A.H. Khandokerl et al., [3] demonstrated the effectiveness             performance with respect to the published major gait
of wavelet based multi scale correlation exponents of MFC as              recognition approaches. This paper presents a methodical and
features for automated screening an individual subject of                 comprehensive gait recognition approach, which can work just
proper balance control as being within healthy ranges, or                 as fine as other complex published techniques in terms of
having high enough risk to be categorized as a falls risk or a            effectiveness of performance while providing all the
faller by using SVM. Findings of that study were based on a               advantages associated with the computational efficiency for
small sample of impaired subjects, compared to a relatively               real-world applications.
small sample of healthy peers. Therefore, further validation of              Shakhnarovich et al. combined the face and MV-based gait.
the relative risk estimation task is suggested in a larger, more          The front face was captured by one camera and the side-view
diverse sample of healthy and balance impaired falls risk in              of the person was captured by another camera. Face-alone,
elderly adults, which may subsequently lead us to make more               gait-alone and combined face and gait recognition rates were
robust automated diagnostic model of falls risk estimation.               80%, 87%, and 91%, respectively. Zhou et al. [21] used a
The significance of this study is that it provides an early               single camera to capture both face and gait. Recognition rates
estimation of relative falls risk in the elderly that holds great         for face and gait separately were 64.3% and 85.7%, a single
potential for indicating balance improving interventions to               respectively. Conversely, when they were combined, the
reduce their relative risk of falls.                                      recognition rate increased up to 100% [23]. In [22], WS-based
   Jian Ni et al., [17] had a look about gait recognition, which          gait recognition was combined with speaker verification.
is simulated in the small and medium-scale gait database.                 Performance proved to be appreciably better in a noisy
Higher recognition rate and faster recognition speed of the               environment, compared to when speaker verification was used
algorithm are verified. The reason that this algorithm obtains            alone. The EER was in the range of 2%-12%, less than half of
superior test results is: The paper adopts support vector                 the EER of individual modalities. In this group, gait is
machine based on hybrid kernel function. This method makes                captured using a video-camera from distance. Video and
that the SVM model has better generalization ability. In the              image processing techniques are employed to extract gait
method of parameter selection, the text uses the objective                features for recognition purposes.
function and combines OPS algorithm to select the best kernel                BenAbdelkader et al. [25] used stride and cadence for
parameter. The way combines the advantages of objective                   person identification and verification. Johnson and Bobick
function and PSO algorithm to optimize SVM parameters. It                 [26] extracted static body parameters such as the height, the
significantly improves the optimization speed, at the same                distance between head and pelvis, the maximum distance
time obtains a good optimization effect.                                  between pelvis and feet, and the distance between feet, and

                                                                                                     ISSN 1947-5500
                                                                 (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                           Vol. 8, No. 5, 2010
used them for recognition. Most of the MV related gait
recognition algorithms are based on the human silhouette [27,
28]. That is the image background is detached and the
silhouette of the person is extracted and analyzed for
recognition, For example, Liu and Sarkar [27] computed the
average silhouettes over a gait cycle, and used the Euclidean
distance between them to compute similarity.

                      III.   METHODOLOGY
The block diagram of the proposed system is shown in
Figure1. After suitable preprocessing, the Salient Gait
Features are extracted from the possible Gait Signatures.
These Gait features are then subjected to t-test normalization             There are k classes.
and subsequently to SVM Classifier.

Preprocessing            Deriving Gait               Gait Feature
 Gait Images              Signature                   Extraction
                                                                           is the maximum of all k. Ck refers to class k that includes 
                                                                              , samples      is the expression value of feature i in sample j
                                                                                is the mean expression value in class k for feature i. where
SVM Classifier               Feature                   Statistical         n is the total number of samples. xi is the general mean value
   Normal/               Normalization               based Feature         for feature i. Si is pooled within-class standard deviation for
  Abnormal                                             Selection           feature i. In fact, the TS used at this point is a t-statistic
                                                                           between the centroid of a specific class and the overall
                    Figure1: Proposed Architecture                         centroid of all the classes. Another possible model for TS
A.     Statistical Methods for Gait Feature Selection &                    could be a t-statistic between the centroid of a specific class
Normalization                                                              and the centroid of all the other classes.
                                                                              Two samples are given as input to the T-Test. The paired t-
The main objective of feature selection is to discover a subset
                                                                           test determines whether input features differ from each other
of features, satisfying certain criteria. In pattern recognition,
                                                                           in a significant way under the assumptions that the paired
recognition metric will be the classification accuracy or
                                                                           differences are independent and identically normally
inversely the classification error. But direct minimization of
                                                                           distributed. This gives a clear view for Abnormal Gait
the classification error cannot be analytically performed, so a
wide range of alternative statistics that are easier to evaluate
are performed. The typical measure used in the gait feature
                                                                               2.   PCA (Principal Component Analysis)
selection is introduced as follows:
                                                                              Principal Component Analysis (PCA) involves a
     1. T-Test
                                                                           mathematical process that transforms a number of possibly
   The t-test finds whether the means of two groups are
                                                                           correlated variables into a smaller number of uncorrelated
statistically dissimilar from each other. This analysis is
                                                                           variables called principal components. The fallouts of a PCA
appropriate to compare the means of two groups, and
                                                                           are usually discussed in terms of component scores and
especially appropriate for the analysis of the two-group
randomized experimental design.
                                                                              PCA is the simplest of the proper eigenvector-based
                                                                           multivariate analyses. Often, its process can be thought of as
The t-score (TS) [31] of feature i is defined as follows:
                                                                           revealing the internal structure of the data in a way which best
                                                                           explains the variance in the data. If a dataset (multivariate) is
                                                                           visualized as a set of coordinates in a high-dimensional data
                                                                           space (1 axis per variable), PCA supplies the user with a
                               Where,                                      lower-dimensional picture, a "shadow" of this object when
                                                                           viewed from its (in some sense) most informative viewpoint.
                                                                           For a data matrix,     with zero empirical mean (the empirical
                                                                           mean of the distribution has been subtracted from the data set),
                                                                           where each row represents a different repetition of the

                                                                                                       ISSN 1947-5500
                                                                           (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                                     Vol. 8, No. 5, 2010
experiment, and each column gives the results from a                                                       ⎧n
particular probe, the PCA transformation is given by:                                                                     (      )
                                                                                                       sgn ⎨∑ ai∗ y ∗ k x, xi∗ + b}
                                                                                                           ⎩ i =1

                                   =V                                                     C SVM kernel functions
                                                                                     This work is the first attempt to test the classification ability of
                                                                                     feature combinations in gait applications. We have used three
where the m-by-n diagonal matrix Σ is an nonnegative real
                                                                                     main kernel functions for our study here. The Partial kernel
numbers on the diagonal and W Σ VT is the Singular Value
Decomposition (SVD) which is factorization of a rectangular                          function, which only influence to data near the test points and
real or complex matrix of X.                                                         the kernel function more applied here is the Radial Basis

   B      Support Vector Machine                                                                                      {
                                                                                                     k (x, x k ) = exp − x − x k
                                                                                                                                         2δ 2 ,
                                                                                       where δ is the width of the gaussian kernel. The overall
The support vector machine is used as a classifier in the paper.
SVM is the one of the best linear classification method and                          kernel function which allows that, the data away from the test
kernel mixed applications. The SVM transforms the samples                            points will also have impact to kernel function. It is the
to high-dimension space by the kernel mapping, and then get                          polynomial kernel to be more suitable in this case:
the best linear classification surface of samples in this new
space. This Non-linear transformation is achieved by                                                         k(x, xk)=[(x, xk)+1]d,,
appropriate inner product function. The best linear
classification surface function of characteristics space can be                      where d is degree of polynomial.
described by the formula:
                                                                                                     IV.      EXPERIMENTAL RESULTS
                  g ( x ) = Σ a j y i k ( x, xi ) + b

                               j =i                                                     This study mainly deals with the performance analysis of
                                                                                     the T-Test based SVM classification method for gait normality
   Where (xi, yi) are the two types of sample collection divided                     and abnormality. In this section, several experiments are
in the sample space, b is the classification threshold, and k(x,                     carried out to test the validity of T-Test based SVM. A
xi) is being the nonlinear kernel function that replaces                             comparative analysis is also done for the proposed T-Test with
characteristics space and meets Mercer conditions. Ascertain                         PCA (Principal Component Analysis). The experimental data
the best linear classification surface function is got by striking                   used in this study are obtained from the gait database of
the best resolve ai where i = 1, 2,…,n of the following function                     Virginia University [11]. There are totally 158 gait samples
Q(a).                                                                                present in the database and all these samples are used for this
                                                                                     experiment. These samples belong to 68 children with normal
                          g (x ) = Σ a j y i k

                                          j =i                                       gait and 88 children with abnormal gait affected with cerebral
                                                                                     palsy (CP). The ages of these children range from 2 years to
                                                                                     13 years. Four features of gait samples are selected for
                               − 0 .5 ∑ ∑ a i a j y i y j k (x i , x j )
                     n                        n   n
     max Q (a ) =   ∑a     i                                                         classification and they are stride length, cadence, leg length
                    i=0                    i=0 j =0                                  and age.
                                                                                        In this study, the t-test is applied to normalize the gait
                                ∑yai =1
                                           i i    = 0,    i = 1,2,..., n             samples. Figure 2 shows the distribution of samples before and
                                                                                     after normalization. As shown in Figure 2, the overlap of two
                                      0 ≤ ai                                         sample sets is effectively reduced after normalization, which
                                                                                     helps to improve the classification accuracy. Three kernel
  The above equation is solving of quadratic function extreme                        functions are used to build SVM classifiers in this study. By
value on condition that inequality, Q (a)           is convex                        comparing the classification results of three classifiers, the
function. Because its local optimal solution is global optimal                       most suitable kernel function may be decided for t-test-SVM.
solution, the solution is unique. Thus the best classification                       They are Radial Basis Function (RBF), linear and the
function of SVM is:                                                                  polynomial. The RBF has best accuracy rate when compared
                                     ⎧n                                              to the other kernels such as the linear and the polynomial. In
      f ( x ) = sgn ( g ( x )) = sgn ⎨∑ a j y j k (x, xi ) + b}                      general, the RBF kernel is a reasonable first choice.
                                     ⎩ j =1

                                                                                                                  ISSN 1947-5500
                                                                      (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                                Vol. 8, No. 5, 2010

                                   (a)                                                                           (b)

       Figure 2: (a) Distribution of Raw Data Before Normalization.                        (b) Distribution of Raw Data After Normalization.

   This kernel nonlinearly maps samples of interest in a higher                 So, the linear kernel order d is set to be 1 in the following
dimensional space so that, unlike the linear kernel, it can                     experiments. When d is 1, the polynomial kernel function
handle the case when the relation between class labels and                      actually is a linear kernel function with accuracy of 90.12%.
attributes is nonlinear. Furthermore, the linear kernel is a                    But in case of the RBF kernel with σ = 4, c = 100,the absolute
special case of the RBF since the linear kernel with a penalty                  accuracy rate is of 98.15%, which leads the other two kernels.
parameter C has the same performance as that of RBF kernel                      When radial basis function (RBF) is applied, the kernel
with some parameters. In addition, the sigmoid kernel behaves                   parameter σ in the RBF and the regularization parameter C
similar to the RBF for certain parameters. The next reason is
the number of hyper parameters that influences the complexity                   may impact the classification accuracy of T-Test-SVM. Figure
of model selection. The polynomial kernel has additional                        4 elaborately shows the relationship between classification
hyper parameters than that of the RBF kernel.                                   accuracy and parameters combination (d, C).
   Finally, the RBF kernel has fewer numerical difficulties.
One key point should lie between 0 and 1, in contrast to
                                                                                        TABLE 1: BEST ACCURACY ACHIEVED FOR KERNEL
polynomial kernels and linear kernel of which kernel values                                             FUNCTIONS
may go to infinity or zero, while the degree is large. Moreover,
that the sigmoid kernel is not valid product of two vectors,
under some parameters and conditions.
   There are a few situations where the RBF kernel is not                          Kernel Function          Parameters               Accuracy (%)
suitable. In meticulous, when the number of features is very                           Linear                  d=1                      90.12
large, one may just use the linear kernel.                                           Polynomial                d=1                      85.69
                                                                                        RBF                σ=4, C=100                   98.15
By applying this kernel function, the accuracy of t-test -SVM
and PCA SVM are compared at the end of this section.                            The classification accuracy of the RBF kernel function rates
Polynomial order d is an important parameter when                               high of order 98%, this is shown in figure 4 for all the three
polynomial kernel function is applied in t-test-SVM. The                        cases of C=1, C=10, C=100.
classification accuracy of gait samples by using different                         The Figure reflects that, the generalization capability of the
polynomial order is shown in Figure 3, and d is chosen from 1                   SVM enhances along with the increase of C. This is because
to 10. As shown in Figure 3, the classification accuracy                        the regularization parameter C may adjust the ratio of
declines along with the increase of polynomial order. This is                   confidence interval and empirical risk.
since the aspect of the feature space is high under a large
polynomial order and it leads a declining generalization
capability of SVM.

                                                                                                             ISSN 1947-5500
                                                                 (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                           Vol. 8, No. 5, 2010
                                                                           These analysis reveals that T-Test have accuracy rate of
                                                                           98.15%, while that of the PCA is comparably low.

        Figure 3, Classification Accuracy and Polynomial Order

                                                                                 Figure 4, Classification Accuracy for Various Parameter Combinations

      σ              C=1              C=10              C=100                                           V.      CONCLUSION
      0             20.12             54.21             92.07              In this paper Abnormal Gait Analysis is done using the SVM
      1             35.08             63.75             94.29              with T-Test Combination. The results are compared with the
      2             54.82             78.64             94.38              other existing methods based on their Classification accuracy.
      3             65.54             85.11             95.93              An automated abnormal gait classification system is described
                                                                           using machine learning techniques. To achieve this, the gait
      4             75.12             90.39             96.17
                                                                           signature has been extracted by combining a statistical
      5             90.43             92.42             97.59              approach and machine learning based analysis is further done
      6             92.54             92.78             97.26              using the anatomical knowledge. For the derived gait
      7             93.72             96.64             98.07              signatures, the motion parameters were calculated, and the gait
      8             92.05             96.89             98.15              features based on the motion parameters were extracted. The
                                                                           T-Test based SVM classifier is used to analyze the
      9             92.34             96.92             98.15
                                                                           discriminatory ability of the extracted features. The result of
      10            92.16             96.11             98.05              the proposed method has produced very good classification
                                                                           rate which exceeds 98%. As such, the automated abnormal
   In this case, the SVM has almost no change of the empirical             classification system not only accords with quantitative
risk and generalization capability. Table I shows the                      analysis in results, but also confirms distinctiveness as normal
classification accuracy of the three classifiers.                          and abnormal gait. Hence this gait classification for medical
                                                                           diagnosis would be a real boon as its convenience will surely
  TABLE 3: BEST ACCURACY ACHIEVED FOR VARIOUS METHODS                      benefit the children and also the elderly. The drastic
                                                                           development of computer vision techniques also ensures that
             Algorithm                      Accuracy (%)                   the clinical gait analysis put into practical realization may
             Std SVM                           87.68                       gradually be achieved.
            PCA SVM                            96.51                                                       REFERENCES
            T-Test SVM                         98.15                       [1]    Mostayed, A. Mynuddin, M. Mazumder, G. Sikyung Kim Se Jin Park,"
                                                                                  Abnormal Gait Detection Using Discrete Fourier Transform",
                                                                                  International Conference on Multimedia and Ubiquitous Engineering,
The generalization capability of the SVM is weak when C is                        2008.
small, because a small C indicates that the punishment for                 [2]    Bauckhage, C. Tsotsos, J.K. Bunn, F.E.,"Detecting Abnormal Gait",
empirical risk is small and hence the possibility of the                          Proceedings of Canadian Conference on Computer and Robot Vision,
empirical risk is large. When C exceeds a certain value, the                      2005.
complexity of a classifier reaches the maximum allowed limit               [3]    A.H. Khandokerl, Daniel Lai, Rezaul K. Begg and Marimuthu
                                                                                  Palaniswami,"A Wavelet Based Approach for Screening Falls Risk in
in the feature space.                                                             the Elderly using Support Vector Machines", Fourth International
   Table 3 shows, that the T-Test SVM algorithm have better                       Conference on Intelligent Sensing and Information Processing, 2006.
accuracy rate while compared with other algorithms. The                           ICISIP 2006.
accuracy rate of the PCA and T-Test SVM are compared with                  [4]    Inman, V. T., Ralston, H. J., and Todd, F.: Human Walking. Williams &
                                                                                  Wilkins, Baltimore (1981).
the different parameters for a particular dataset of children.

                                                                                                             ISSN 1947-5500
                                                                           (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                                     Vol. 8, No. 5, 2010
[5]    Murray, M. P., Drought, A. B., and Kory, R. C.: Walking Patterns of              [25] C. BenAbdelkader, R. Cutler, and L. Davis. Stride and cadence as a
       Normal Men. Journal of Bone and Joint Surgery, 46A(2) (1964) 335-                     biometric in automatic person identification and verification. In Fifth
       360.                                                                                  IEEE International Conference on Automatic Face and Gesture
[6]    Johansson, G.: Visual Perception of Biological Motion and a Model for                 Recognition, pages 357–362, May 2002.
       Its Analysis. Perception and Psychophysics, 14(2) (1973) 201-211.                [26] Amos Y. Johnson and Aaron F. Bobick. A multi-view method for gait
[7]    Winter, D. A.: The Biomechanics and Motor Control of Human Gait:                      recognition using static body parameters. In Third International
       Normal, Elderly and Pathological. Waterloo Biomechanics, Ontario                      Conference on Audio- and Video-Based Biometric Person
       (1991).                                                                               Authentication, pages 301–311, June 2001.
[8]    Kyoungchul Kong Tomizuka, M," A Gait Monitoring System Based on                  [27] Zongyi Liu and Sudeep Sarkar. Simplest representation yet for gait
       Air Pressure Sensors Embedded in a Shoe", IEEE/ASME Transactions                      recognition: Averaged silhouette. In International Conference on Pattern
       on Mechatronics, 2009.                                                                Recognition, pages 211–214, 2004.
[9]    Nixon, M. S., Cater, J. N., Grant, M. G., Gordon, L., and Hayfron-               [28] Zongyi Liu, Laura Malave, and Sudeep Sarkar. Studies on silhouette
       Acquah, J. B.: Automatic Recognition by Gait: Progress and Prospects.                 quality and gait recognition. In Computer Vision and Pattern
       Sensor Review, 23(4) (2003) 323-331.                                                  Recognition, pages 704–711, 2004.
[10]   J. Devore and R. Peck, Statistics: The Exploration and Analysis of Data,         [29] Yanmei Chai, Jinchang Ren, Rongchun Zhao, and Jingping Jia.
       third ed. Duxbury Press, 1997.                                                        Automatic gait recognition using dynamic variance features. In
                                                                                             International Conference on Automatic Face and Gesture Recognition,
[11]   IEEE transactions on rehabilitation engineering, vol. 5, no. 4, december              pages 475 – 480, 2006.
[12]   Foster, J. P., Nixon, M. S., and Prügel-Bennett, A.: Automatic Gait                   common.htm.
       Recognition using Area-based Metrics. Pattern Recognition Letters,
       24(14) (2003) 2489-2497.                                                         [31] M. Pushpa Rani and G.Arumugam, “An Efficient Gait Recognition
                                                                                             System For Human Identification Using Modified ICA “,International
[13]   Christian Bauckhage, John K. Tsotsos, Frank E. Bunn, "Automatic
                                                                                             Journal of Computer Science & Information Technology, vol. 2, No 1,
       Detection of Abnormal Gait", International journal of Image and Vision                pp 55-67,January 2010.
       Computing, 2007.
[14]   Zongyi Liu and Sudeep Sarkar, "Improved Gait Recognition by Gait
       Dynamics Normalization," IEEE Trans on Pattern Analysis and Machine                                          AUTHORS PROFILE
       Intelligence, Vol. 28, No. 6, June 2006.
[15]   L. Lee, G. Dalley, and K. Tieu, “Learning Pedestrian Models for                       Ms.M.Pushpa Rani is an Associate Professor in Computer Science at
       Silhouette Refinement,” Proc. Int’l Conf. Computer Vision, pp. 663-                                          Mother Teresa Women’s University, Tamil
       670, 2003.                                                                                                   Nadu, India. She received her Master’s
[16]   A. Veeraraghavan, A.R. Chowdhury, and R. Chellappa, “Matching                                                degree in Computer Applications from
       Shape Sequences in Video with Applications in Human Movement                                                 Bharathiar University, Coimbatore, India,
       Analysis,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol.                                       and is currently pursuing doctoral Research
       27, no. 12, pp. 1896-1909, Dec. 2005.                                                                        in Madurai Kamaraj University, Madurai,
                                                                                                                    India. She has published many articles in
[17]   Jian Ni, Libo Liang, " Gait Recognition Method Based on Hybrid Kernel
                                                                                                                    International Journals and more than 30
       and Optimized Parameter SVM", IEEE International Conference on
                                                                                                                    papers in National and International
       Computer Science and Information Technology, 2009. ICCSIT 2009.
                                                                                                                    Conferences. Her current research areas
[18]   Jang-Hee Yoo, Doosung Hwang," Automated Human Recognition by                                                 include Image Processing, Biometrics and
       Gait using Neural Network", IEEE Conference on Image Processing                                              Adaptive Learning System.
       Theory, Tools & Applications, 2008.
[19]   Ju Han, Bir Bhanu, "Individual Recognition Using Gait Energy Image",
                                                                                             Dr G Arumugam is a senior Professor in the Department of Computer
       IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol.
                                                                                             Science, Madurai Kamaraj University, Madurai, He did his Masters
       28, No. 2, February 2006.
                                                                                             degree in Applied Mathematics specializing in Computer Science in the
[20]   G. Shakhnarovich, L. Lee, and T. Darrell. Integrated face and gait                    PSG College of Technology, Coimbatore in the year 1980. He started
       recognition from multiple views. In Proceedings of the IEEE Computer                  his research carrier from the Department of Mathematics, IIT Kanpur in
       Society Conference on Computer Vision and Pattern Recognition., 2001.                 1981.He got his Doctorate degree from the University of Pierre and
[21]   Xiaoli Zhou, Bir Bhanu, and Ju Han. Human recognition at a distance in                Marie Curie , France in 1987.He was a Post-Doctoral Fellow in the
       video by integrating face profile and gait. In 5th International                      University of JYVASKYLA, Finland for a period of three months.
       Conference on Audio- and Video-Based Biometric Person                                                              He worked in the Hexaware Info Systems,
       Authentication, pages 533–543, July 2005.                                                                          Chennai as a Project Manager and as a
[22]   Elena Vildjiounaite, Satu-Marja Makela, Mikko Lindholm, Reima                                                      Consultant in the Polaris Software, Chennai
       Riihimaki, Vesa Kyllonen, Jani Mantyjarvi, and Heikki Ailisto.                                                     to gain industrial experience.. He was in
       Unobtrusive multimodal biometrics for ensuring privacy and                                                         Singapore for a period of 2 ½ years as
       information security with personal devices. In Pervasive, pages 187–                                               Visiting Professor in the Ngee Ann
       201, May 2006. Springer LNCS.                                                                                      Polytechnic, Singapore from July 2000 to
[23]   G. Shakhnarovich, L. Lee, and T. Darrell. Integrated face and gait                                                 Nov.2002.He has published several papers
       recognition from multiple views. In Proceedings of the IEEE Computer                                               in the international and national journals and
       Society Conference on Computer Vision and Pattern Recognition. 2001.                                               in the conference proceedings. His area of
                                                                                                                          research is design and analysis of
[24]   Davrondzhon Gafurov, "A Survey of Biometric Gait Recognition:                                                      algorithms, data mining, and cryptography
       Approaches, Security and Challenges", NIK-2007 conference.                                                         and network security.

                                                                                                                         ISSN 1947-5500

To top