I. INTRODUCTION:

Biometric systems for human identification at distance have ever been an increasing demand in various significant applications. Many biometric resources, for instance iris, fingerprint, palmprint, hand geometry have been systematically studied and employed in many systems. In spite of their widespread applications, these resources suffer from two main disadvantages: 1) Failure to match in low resolution images, pictures taken at a distance and 2) Necessitates user cooperation for accurate results [4]. For these reasons, innovative biometric recognition methods for human identification at a distance have been an urgent need for surveillance applications and gained immense attention among the computer vision community researchers in recent years. In this modern era, the integration of human motion analysis and biometrics has fascinated several security-sensitive environments such as military, banks, parks and airports etc and has turned out to be a popular research direction.

Human gait recognition works from the observation that an individual’s walking style is unique and can be used for human identification. So as to recognize individual’s walking characteristics, gait recognition includes visual cue extraction as well as classification. But the major issue here is the representation of the gait features in an efficient manner. Two common categories of gait recognition are appearance-based and model-based approaches.

Among the two, the appearance-based approaches suffer from changes in the appearance owing to the change of the viewing or walking directions [13]. But, model-based approaches extract the motion of the human body by means of fitting their models to the input images. Modelbased ones are view and scale invariant and reflect in the kinematic characteristics of walking manner [7]. In general, a gait is considered as being composed of a sequence of kinematic characteristics of human (i.e. human motion) and most systems in existence recognize it by the similarity of these characteristics. Compared with those traditional biometric features, such as face, iris, palm print and fingerprint, Gait has many unique advantages such as non-contact, non-invasive and perceivable at a distance. The introduction of gait has turned video-based intelligent security surveillance system [8] as a technology for the future [3].But, gait features have a high intra-personal variation in shape and also it is influenced by external conditions like footwear, clothing and load carrying.

The variation of gaits is also influenced by mood, ground surface condition and time difference [12]. In spite of its individual pros and cons, gait recognition can be thought of as an effective means for human identification at a distance.

Recognition of an individual is an important task to identify people. Identification through biometric is a better way because it associate with individual not with information passing from one place to another. Biometrics is a physiological or behavioral characteristic, which can be used to identify and verify the identity of an individual. There are numerous biometric measures which can be used to help derive an individual identity. They are physiological, like fingerprints, face recognition, iris-scans and hand scans and behavioral, like keystroke-scan and speech patterns. Gait recognition is relatively new biometric identification technology which aims to identify people at a distance by the way they walk. It has the advantage of being unobtrusive, difficult to conceal, noninvasive and effective from a distance. Human gait recognition as a new biometric aimed to recognize person via the style of people walking, which contain the physiological or behavioral characteristics of human.

In [1], the literature describes the general method and development actuality of gait recognition, they describe three methods of gait recognition, which include statistical based method, model based method and fusion based method. The statistical based method characterizes body movement by the statistic of the space temporal pattern generated in the image sequence by the locomotive person. The advantages of this method are low computational cost and less time consuming. Model based method constructs human model to recover explicit features describing gait dynamics such as stride dimensions and the kinematics, of joint angle. The advantage of this method is the ability to drive gait signature from model parameter and free from the effect of the different clothing and view point. However, it is time consuming and costly [5]. Fusion is combination of both statistical and model based method.

II. PREVIOUS WORK:

The problem of image-based human motion analysis and recognition has been receiving considerable attention in the iterature. Most of the proposed approaches involve tracking the pose of the human body, represented either as kinematic chain of body parts [11, 15, 12], or as spatial arrangement of blobs [9] or point features [1]. Statistical models, such as standard [5, 1] and parametric [6, 10] Hidden Markov Models are then fitted to the tracking data and likelihood tests are used for recognition. In [10, 12, 2] mixed state statistical models for the representation of motion have been proposed, and in [18, 19] particle filters have been applied in this framework for estimation and recognition. In

[4] linear gaussian models have been used, and recognition is performed by defining a metric on the space of models. Other techniques do not require an explicit model of the human body. Zelnik-Manor et al. [15] propose a statistics of the spatiotemporal gradient at multiple temporal scales and use it to define a distance between video sequences. Some approaches [14, 9, 8] are specific to recognition of periodic motion, such as the human gaits we consider in this paper. In [14] classification is based on periodicities of a similarity measure computed on tracked moving parts. Little and Boyd [8] use Fourier analysis to compute the relative phase of a set of features derived from moments of optical flow, and employ the resulting phase vector for classification. Bobick and Davis [7] propose a description based on the spatial distribution of motion, the Motion Energy and Motion Histogram Images. Recognition is done by comparing Hu moments [13] of those images with a set of stored models. In [12], the problem is recognizing actions from video taken from a distance, where the person appears only as a small patch. They compute a set of spatiotemporal motion descriptors on a stabilized figure-centric sequence, and match the descriptors to a database of preclassified actions using nearest neighbor classification.

Given the ability of humans to identify persons and classify gender by the gait of a walking subject, there have been a few computer vision algorithms developed for people identification and activity classification. Cutler and Davis [2] used self-correlation of moving foreground objects to distinguish walking humans from other moving objects such as cars. Polana and Nelson[6] detected periodicity in optical flow and used these to recognize activities such as frogs jumping and human walking. Bobick[1] used a time delayed motion template to classify activities. Little and Boyd[5] used moment features and periodicity of foreground silhouettes and optical flow to identify walkers.

Figure 1: An example sequence of a walking person

Han and Bhanu [3] use gait energy image for gait analysis. Tey used statistical feature extraction approach for learning effective feature and feature fusion strategy is used to improve recognition. In [4], Eigenspace transformation based on Principal Component Analysis (PCA) is applied to reduce the dimensionality of the input feature space. Then supervised pattern classification techniques are finally performed in the lower-dimensional eigenspace for recognition. Su and Zanga [5] use fuzzy principal component for recognition. Firstly they processed the original gait sequence and gait energy image is obtained then Eigen value and Eigen vector are extracted by fuzzy principal component analysis, which are called fuzzy logic. Finally NN classier is utilized in feature classification. In [6], proposed low resolution method used manifold sampling, back projection and multi linear tensor based learning without tuning parameter. Davrondzhon proposed important gait recognition using cycle matching in which they use wearable accelometer, to record ankle motion for measuring cycle [7].

In [17], presents an original 3D approach for automatic gait recognition based on analyzing image sequences captured by stereo vision. Contour matching is done after binarized silhouette of a moving individual is firstly achieved in order to get 3D contour. Then, stereo gait feature (SGF) which is the norm of stereo silhouette vector (SSV) is extracted from 3D contour Principal Component Analysis (PCA) is adopted for dimensionality reduction. Finally, NN and ENN is applied for classifying and distinguishing.

In order for the biometrics to be ultra secure and to provide more than average accuracy more than one form of biometrics requires. Hence the need arises for the use of multimodal biometrics. This uses a combination of different biometrics recognition technologies. Multimodal biometrics technology uses more than one biometric identifier to compare the identity of the person. When designing a multimodal biometrics system, two factors should be considered: (a) the choice of biometric traits; (b) the level at which information should be fused [9].

III. PROPOSED APPROACH:

Figure 2: Proposed System

In this research, we propose an efficient human gait recognition system using modified Independent Component Analysis (MICA). The proposed gait recognition system characterizes gait in terms of a gait signature computed directly from the sequence of silhouettes. The system can be seen as a generic pattern recognizer composed of the three main modules namely, i) Human detection and tracking ii) Training using Modified ICA and iii) Human recognition. Initially, the moving objects (human) are segmented and tracked in each frame of the given video sequence (tracking module). Then, the person’s identity is determined by training and testing using MICA on the extracted feature vectors (pattern recognition module). Fig.2. depicts the block diagram of the Proposed Gait Recognition System

3.1. Human Detection and Tracking

Detection and tracking of human from a video sequence is the first step in gait recognition. The system works with the assumption that the video sequence to be processed is captured by a static camera, and the only moving object in video sequence is the subject (person). Given a video sequence from a static camera, this module detects and tracks the moving silhouettes. This process comprises of two submodules:1) Foreground Modelling and 2) Human tracking using skeletonization operation.

3.1.1. Foreground Modelling

Background subtraction has been extensively used in foreground detection, where a fixed camera is usually used to capture dynamic scenes. To reliably generate the background image from video sequences is critical [5]. In the proposed system a simple motion detection method based on median value is adopted to model the background from the video sequence.

Let P represent a video sequence having N image frames. The background P(x, y) can be constructed using the formula:

( , ) [ ( , ), ( , ),.......... ( , )] 1 2 P x y median P x y P x y P x y N = (1)

The value of P(x, y) is the background brightness to be calculated in the pixel location (x, y) and median symbolizes its median value. In the proposed gait recognition system, we have computed the median value rather than mean value of pixel intensities over N frames, since,

1) Distortion of the mean value for a large change in pixel intensities while the person moves. The median is impervious to spurious values and

2) Median value Computation is comparatively faster than the least mean square value [5]. Both these statements hold with the assumption that a person continuously moves around over the frames [24].

Subsequently, the extracted background and the original image frames are provided for the foreground modelling. The background subtraction algorithm subtracts the background from the original image frames to obtain the moving foreground objects i.e. human subject in binary.

3.1.2. Human Tracking

The next step is to the track the moving silhouettes of a walking figure from the extracted binary foreground image. We adopt the morphological skeleton operator for human tracking. Skeletonization is defined as the process for reducing foreground regions in a binary image to a skeletal remnant that greatly preserves the degree and connectivity of the original region while removing a good number of the original foreground pixels. If A is a set in the plane, then the disk rD_Xof radius r and centered at x is maximal with respect to A if it is contained in A and is not properly contained in any other disk contained in A. Blum [6] defines the skeleton (medial axis) of A, denoted by SK(A) , to be a set of centers of all disks maximal with respect to A. Each point x in the skeleton corresponds to the radius of the maximal disk centered at x . The skeleton function S(x) returns the radius of the maximal disk centered at x . The rth skeleton subset of SK(A) , denoted by S (A) r , is defined to be the set of all skeleton points x such that S(x) = r . An important property of the skeleton function is that, it contains all necessary information to reconstruct the original image, i.e.

The above fact has been made use of to efficiently encode binary images by their skeletons [27]. Lantuejoul [28] has shown that the skeleton can be defined by means of morphological Operations

Where, rD denotes the open disk of radius r , drD is the closed disk of infinitely small radius dr and “-“ denotes set difference [8].

3.2. Training using MICA

We use the modified Independent Component Analysis (MICA) to extract and train the gait features. The purpose of training the skeletonised silhouettes with the modified ICA is to attain a number of independent components to represent the original gait features from a high dimensional measurement space to a low-dimensional Eigenspace. The concept of ICA can be noticed as a generational of Principal Component Analysis (PCA) and its fundamental idea is to symbolize a set of random variables using basic functions, where the components are statistically independent or as independent as possible [5]. ICA aims to identify the vectors that describe data to its best in terms of reproducibility; nevertheless these vectors may not comprise of any effective information for classification, and may eliminate discriminative information [10]. The training process is illustrated as follows:

Let { , , ,......, } 1 2 3 n x x x x be the N samples from L classes { , , ,......, } 1 2 3 L w w w w for training and each class represents a sequence of distance signals of one subject’s gait and p(x) their mixture distribution. In a sequel, it is assumed that a priori probabilities ( ) i P w , i = 1,2,...,L , are known. Let m and _ denote the mean vector and the covariance matrix of samples, respectively. ICA algorithm can be used to find a subspace whose basis vectors correspond to the maximum variance directions in the original n dimensional space.

3.3. Human Recognition

With the trained MICA in hand, the final step is to test the effectiveness of the proposed system for gait recognition. Gait recognition has been a traditional pattern classification problem which can be solved by calculating the similarities between instances in the training database and the test database. Gait can be described as a kind of spatiotemporal motion pattern; hence we transform the input gait video sequence into an equivalent parametric eigenspace using the modified ICA (section 3.2.1). Then, based on the similarity measurement computed between the reference patterns and test sample in the parametric eigenspace, we achieve gait recognition. To be more particular, we have used the L2 Norm Distance for measuring the similarity between two gaits. The L2 Norm Distance measure is calculated as

IV. RESULT:

4.1. Data Acquisition: The experimentation of the proposed gait recognition system is performed with images publicly available in the National Laboratory of Pattern Recognition (NLPR) gait database. A brief description of the gait database taken for study: A digital camera (Panasonic NV-DX100EN) fixed on the tripod was used for capturing gait sequences in an open-air environment. The images correspond to a single subject poignant in the field of view without occlusion.

Figure 3. Sample image sequences in the NLPR gait database

The subjects were asked to walk to a stationary camera frontally, laterally, and obliquely (0o, 45o, 90o with respect to the image plane) respectively. The resulting Chinese National Laboratory of Pattern Recognition (NLPR) gait database [6] included 20 subjects and four sequences per view per subject. The properties of the images are: 24-bit full colour, capturing rate of 25 frames per second and the original resolution is 352 x 240. The database comprises a total of 240 sequences. The length of each sequence varies with the time each person takes to traverse the field of view. Some of those image samples are shown in Fig. 2.

4.2. Results

This subsection contains the results of the experiments. The publicly available NLPR gait database is employed in training the MICA. The intermediate results of the presented gait recognition system are depicted in Figure 3, Figure 4, Figure 5 and Figure 6 respectively.

Figure 4. Sample images of a subject in the database

Figure 5. Extracted Background

Figure 6. Extracted silhouettes of frames in Figure. 4

Figure7. Human tracked by skeletonization

Figure We have evaluated the effectiveness of the proposed system with a set of gait images available in the NLPR database. Moreover, we have measured the False Acceptance Rate (FAR) and False Rejection Rate (FRR) corresponding to the proposed gait recognition system and presented in Table. 1.

V. CONCLUSION:

With mounting demands for visual surveillance systems, human identification at a distance has recently emerged as an area of significant interest. Gait is being considered as an impending behavioural feature and many allied studies have illustrated that it can be used as a valuable biometric feature for human recognition. The development of computer vision techniques has also assured that vision based automatic gait analysis can be gradually achieved. The proposed system has been tested on the gait databases and, the extensive experimental results on outdoor image sequences demonstrated that the proposed system possesses a pleasing recognition performance.

VI. REFERENCES:

[1] D. Gavrila, “The Visual Analysis of Human Movement: A Survey,” Computer Vision and Image Understanding, Vol. 73, No. 1, pp. 82-98, 1999.

[2] L. Wang, W.M. Hu, and T.N. Tan, “Recent Developments in Human Motion Analysis”, Pattern Recognition, Vol. 36, No. 3, pp. 585-601, 2003.

[3] Jiwen Lu, Erhu Zhang, "Gait recognition for human identification based on ICA and fuzzy SVM through multiple views fusion", Pattern Recognition Letters, Vol. 28, pp. 2401–2411, 2007.

[4] Dacheng Tao, Xuelong Li, Xindong Wu, and Stephen J. Maybank, "General Tensor Discriminant Analysis and Gabor Features for Gait Recognition", IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 29, No. 10, pp. 1700-1715, October 2007.

[5] Liang Wang, Tieniu Tan, Huazhong Ning, and Weiming Hu, "Silhouette Analysis-Based Gait Recognition for Human Identification", IEEE transactions on pattern analysis and machine intelligence, Vol. 25, No. 12, December 2003.

[6]“Gait Database” from http://www.cbsr.ia.ac.cn/english/ IrisDatabases.asp

[7] X. Yang, Y. Zhou, T. Zhang, G. Shu, J. Yang, “Gait Recognition Based on Dynamic Region Analysis,” Signal Processing, Vol. 88, No. 9, pp. 2350–2356, September 2008.

[8] W.M. Hu, T.N. Tan, L. Wang, S. Maybank, “A survey of visual surveillance of object motion and behaviours”, IEEE Trans. Syst. Man Cybern., Part C, Vol. 34, No. 3, pp. 334–352, 2004.

[9] Jianyi Liu, Nanning Zheng, and Lei Xiong, "Silhouette quality quantification for gait sequence analysis and recognition", Signal Processing, Vol. 89, No. 7, pp. 1417-1427, July 2009.

[10] Xiaoli Zhou and Bir Bhanu, "Feature fusion of side face and gait for video-based human identification", Pattern Recognition, Part Special issue: Feature Generation and Machine Learning for Robust Multimodal Biometrics, Vol. 41, No. 3, pp. 778-795, March 2008.

[11] Changhong Chen, Jimin Liang, Heng Zhao, Haihong Hu and Jie Tian, "Frame difference energy image for gait recognition with incomplete silhouettes", Pattern Recognition Letters, Vol. 30, No. 11, pp. 977-984, 1 August 2009.

[12] Seungkyu Lee, Yanxi Liu, Collins, R., "Shape Variation-Based Frieze Pattern for Robust Gait Recognition", IEEE Conference on Computer Vision and Pattern Recognition, CVPR '07, pp. 1-8, 17-22

June 2007.

[13] Chan-Su Lee, Ahmed Elgammal, "Gait Style and Gait Content: Bilinear Models for Gait Recognition Using Gait Re-sampling", Proceedings. Sixth IEEE International Conference on Automatic Face and Gesture Recognition, pp. 147-152, May 2004.

[14] R. T. Collins, R. Bross, and J. Shi, “Silhouette–based Human Identification from Body Shape and Gait,” Proc. IEEE Int’l Conf. Automatic Face and Gesture Recognition, Washington DC, pp. 351–356, 2002.

[15] J. Han and B. Bhanu, “Statistical Feature Fusion for Gait–Based Human Recognition,” Proc. IEEEInt’l Conf. Computer Vision and Pattern Recognition, Washington, DC, vol. 2, pp. 842–847, 2004

Received on 08.11.2011 Accepted on 10.12.2011

Int. J. Tech. 1(2): July-Dec. 2011; Page 137-142