Attention-based gait recognition network with novel partial representation PGOFI based on prior motion information

Compared to other recognition tasks, gait recognition has two unique scenarios, i.e., camera-pedestrian angle change scenario and body contour change (e.g., clothing change) scenario. The current gait recognition methods suffer from feature dilution and fail to extract accurate and highly robust gait features, therefore encounter serious performance degradation when facing these scenarios. In this paper, we propose an attention-based gait recognition network with novel gait representation. First, we design a novel partial gait representation: Part-based Gait Optical Flow Image. During the generation of representation, different parts of the body are separated according to their movement patterns and the optical flow of each part is extracted separately. Second, we propose Prior-Information-based Attention Module to highlight gait features of body parts with distinct motion based on prior information. In terms of appearance features, we propose Attention-based Frame Selection Module to acquire and high-light the key frames. These two modules extract and enhance local features in terms of motion and appearance respectively, avoiding unfocused global feature extraction and solving the feature dilution problem. Finally, our network uses a fusion optimization strategy to allow the network to adaptively balance the contributions of the motion feature and appearance feature, enhancing the robustness of the network under multiple angles. Experiments demonstrate that the method proposed in this paper achieves the best performance on both CASIA-B and OU-MVLP datasets.