Gait Recognition and its Application in Biometric

Abstract

Gait recognition is a novel biometric technique which can be used to identify a person based on his movement. Every person shows a different body movement while walking; particularly limb movements. It has several benefits over existing biometric approaches like a fingerprint scanner or face-recognition. It can be completely unobtrusive, i.e. we can take samples from a distance without the consent of the subject, and works with low-resolution videos. There are various steps included in the gait recognition system, namely, preprocessing, feature extraction, classification, comparing results, etc. Preprocessing includes background subtraction for silhouette extraction, normalization and translation of silhouettes so that a generalized system can be designed, etc. Various features were extracted from the obtained silhouettes, such as, gait cycle, gait energy image, skeleton of silhouettes, angles between body parts, various body lengths, etc. While working with feature extraction it was seen that lower part of human body is more significant while determining gait of a person. The lower part is easily available with gait energy image so this dissertation focuses more on gait energy image (GEI) for classification of data. For classification two techniques are used, sparse via orthogonal matching pursuit (OMP) and Linear Discriminant Analysis. The main technique used is linear discriminant analysis (LDA) which best suits for gait recognition due to unavailability of more samples and high dimensionality of data. LDA works best when dimensionality is high and total no. of samples is less, which exactly the case in gait recognition. Gait has many different features which will give it many dimensions, and since gait recognition is mainly used for surveillance, it often has to be done with low samples of videos and low resolution of frames. LDA was performed using these features and Gait Energy Image, which was done for the first time in this paper. Results are found by comparing accuracy and time taken for recognition. Accuracy is measured by taking into consideration False Rejection Ratio (FRR) and False Acceptance Ratio (FAR). . Results were compared against OMP in MATLAB, and it was found that LDA outperforms OMP both, in the time required as well as in the accuracy.