چكيده به لاتين
Gait recognition is an interesting research field in computer vision. Gait recognition means identifying people by the way they walk. In this thesis, two feature templates are introduced for gait recognition. First one is generated by aggregating the responses of convolving spatial and temporal kernels to the input video sequence. Derivative of Gaussian and truncated cosine are used as kernels to extract spatial and temporal information, respectively. Aggregation of responses over the given gait cycles, form a Gait Spatio-temporal Energy model Image (GSTEI). Furthermore, Dual-Tree Complex Wavelet Transform (DT-CWT) is used to generate the second feature template. DT-CWT is applied to gait images in an arbitrary decomposition level and the magnitude of the resulting six band-pass sub-images is computed. The feature template is generated by concatenating these sub-images into a single image. Besides, overlearning is a common problem in gait recognition, which caused by the high dimensionality of the feature space compared to the small number of training samples. To avoid overlearning, and reduce the effect of covariate factors, a classifier ensemble method called Random Subspace Method (RSM) is used for classification. Experimental results on well-known public databases (i.e., USF and CASIA-B) demonstrate the efficiency of the proposed framework and the combination of the developed feature templates with RSM classification. The averaged Rank-1 and Rank-5 Identification rates in the USF dataset are 72.25% and 85.64% for the GSTEI template and 74.48% and 85.21% for the DT-CWT template, respectively.