Action Perception Machine for Gait Recognition

Gait recognition is one of the most promising video-based biometric techniques. Gait is a periodic movement comprising sequential actions that involve different body parts. Therefore, in a gait cycle, complex structural changes in pedestrian silhouettes contain rich temporal semantic information. However, the existing gait recognition methods based on silhouette sequences are significantly insufficient in temporal semantic representation. This study designs an Action Perception Machine (APM) with a temporal attention mechanism to fully utilize temporal gait information and extract salient action features. First, APM adaptively fuse context information of different time scales to perceive diverse actions. Then, the APM realizes cross-time interaction between actions to capture the salient moments. The two folds enable the APM to have strong temporal semantic representation ability for motion sequences. We integrate the APM into a part-based gait recognition framework (APMGait) to extract discriminative fine-grained gait features. Extensive experiments on the CASIA-B, GREW and Gait3D datasets revealed that the APMGait with a simple network outperforms state of the art.