View invariant gait recognition for person re-identification in a multi surveillance camera environment

Abstract

Person re-identification (ReID) across multiple surveillance cameras with non-overlapping fields of view is one of the most significant problems in real-world intelligent video surveillance systems. Due to the unconstrained nature of the problem, gait-based per- son recognition is the only likely identification method to solve the person ReID in this situation. Furthermore, most of the existing ReID algorithms were designed for closed-world scenarios that consider the same descriptors across the camera network re- gardless of the dramatic change in view angle due to different camera positions, which eventually cause them to perform poorly in real-world scenarios. To address this prob- lem, therefore, in this thesis, we present a simple yet effective algorithm for robust gait recognition for person ReID that addresses the challenges that arise from the real-world multi-camera surveillance environment. In this approach, we first designed a novel low- dimensional spatio-temporal feature vector that was extracted from the pose estimation of raw video frames. In this research, we have developed a 50-dimensional feature descriptor by concatenating four different types of spatio-temporal features. These fea- tures are discriminant, and at the same time robust to the variations of different covari- ate factors. Thereafter, a pose sequence having a timestep of length of 28 frames was formed to feed into an RNN-based classifier network. The RNN network consists of two BGRU layers each of which only has 80 GRU cells. The input layer was followed by a batch normalization layer. The output of the recurrent layers was also batch normalized to standardize the activations and finally fed into an output softmax layer. The major- ity voting scheme was employed to process the output to predict the subject ID. For multi-view gait recognition, we also propose a two-stage network in which we initially identify the walking direction from gait video by employing a view angle identification network. Here, the input of the network was a clip of 16 consecutive frames that were preprocessed and resized to 112x112 to feed into a 3D convolutional network based on C3D. The experimental evaluation conducted on two challenging CASIA A and CA- SIA B gait datasets demonstrates that the proposed method has achieved state-of-the-art performance on both single-view and multi-view gait recognition. The experimental re- sult clearly confirms the effectiveness of our proposed approach when compared to the other state-of-the-art methods.