The 9.11 incident deepens the consensus of security enhancement and triggers the need for intelligent visual surveillance systems. The request for smart surveillance motivates much interest in gait recognition, in an attempt to improve security services. This thesis focuses on the problems of gait recognition (especially gait recognition at night time), gait shape features, and how different scales and metrics affect recognition rates. In addition, we present the extension of vector-oriented relevant component analysis to the higher order tensor case. Major contributions of this thesis include the following: * We reduce the impact of self-adjustment in thermal imagery on walker detection using statistical compensation and present a pseudoshape-based night gait recognition framework. For each video sequence, we first preprocess each frame to reduce the effect of halo and self-adjustment and then employ background subtraction to pinpoint moving people with the help of motion constraints derived from consecutive frames. In addition, we use projective pseudoshape features to characterize distinct walkers and exploit linear subspaces to reduce the dimension of the gait features used. Dynamic temporal cues are utilized in the form of equivalence constraints. Experimental results show that the proposed algorithm gives a promising advance in night gait recognition. * We propose a series of vector-based projective shape features based on the observation that ``shape information dominates the performance of gait recognition''. First, normalized silhouette images are projected in four directions (horizontal, vertical, positive diagonal, and negative diagonal) to obtain eight heuristic features. In addition, optimal projective features are obtained through minimizing the reconstruction error in the optimization paradigm. Furthermore, orthogonal projection features are also combined to describe human gait. Experiments demonstrate that this method can produce satisfying recognition performance. * We present noniterative analytic matrix features for gait recognition, based on the fact ``pedestrian images have the property of second order tensor''. First, each silhouette image is normalized to reduce the effect of image size and walker location. Then unilateral partial order tensor product is used to reduce the rank of silhouette images. Additionally, we apply matrix-based discriminant analysis to improve data distribution and favor classification. Four un...
修改评论