Research On Gait Recognition Based On Human Body Motion Analysis

Static and dynamic information of gait can be captured without awareness at a distance. Therefore, gait is one of the most potential biometric features in long distance human identification and gait recognition would be an essential part of intelligent monitoring system. At the beginning, the research status quo of related technology of gait recognition was analyzed. However, it also was pointed out that gait period detection in arbitrary direction had not yet existed any effective method, the backpack problem of gait recognition remains to be resolved, and effective gait characteristics to improve the performance of gait recognition was still a long term goal. In this thesis, gait recognition based on human body motion analysis was deeply studied, especially focusing on gait period detection, expression and extraction of gait features.The problem of gait period detection was transformed to that of regional characteristics analysis in a single frame, namely, the gait period was divided according to region characteristic variable in each frame. A method for gait period detection based on regional characteristics analysis is proposed. Gait video sequences were coarsely classified into frontal and non-frontal ones in the first instance. Swinging arm regions was detected for frontal gaits, instead regional characteristics such as area, centroid, moment, extreme points and bounding box were detected for non-frontal gaits. Not only the computation is small, but also this proposed method has already achieved the precision of human subjective judgment. Especially, gait period detection based on ellipse fitting is robust to noise. Moreover, with the scaling invariance and shift invariance attributes, this method can be used before the standardized and centralized image processing. Therefore, the processing time of earlier work in gait recognition is reduced significantly.As features in one frame were only relevant to those in prior and subsequent frames during walking, a framework for matrix gait recognition based on linear interpolation was proposed. Then, angle projection, Hough transform, Trace transform and Fan-Beam projection were used as instantiation to prove validity of gait recognition framework, which has brought new ideas to solve gait recognition problems. Actually, this gait recognition feature matrix based on linear interpolation which has another different weight energy form was essentially quite different from the normal gait energy image feature.Because backpack problem heavily affect performance of gait recognition, which only changed the local gait shape, a gait recognition algorithm based on subblock matrix was proposed where redundant subblocks were adaptively removed. Subpattern Complete Two Dimensional Principal Component Analysis, Subpattern Complete Two Dimensional Linear Discriminant Principal Component Analysis and Subpattern Complete Two Dimensional Locality Preserving Principal Component Analysis were presented and applied respectively to each effective sub-image to obtain local features. The third one demonstrated an encouraging recognition performance and showed good robustness against backpack changing.Tensor gait recognition method based on linear interpolation was proposed. The number of frames in one gait cycle should be normalized to a certain amount at first. Then one gait sample can be represented as a tensor. Multilinear principal component analysis (MPCA) was employed here for feature extraction, where two kinds of tensor features were vectorized based on variance information and class discrimination information respectively. Then, MPCA combined with Orthogonal Linear Discriminant Analysis (OLDA) gait recognition algorithm was also proposed. In order to further reduce the calculation, the subblock MPCA plus OLDA strategy was adopted, and different blocks were given different weights which embodied importance of distinguishing information. In the CASIA(B) gait database, the recognition accuracy of 99.46% was acquired.