Understanding Dynamic Associations: Gait Recognition via Cross-view Spatiotemporal Aggregation Network

Gait recognition is a technique to recognize subjects under long-distance conditions. Existing gait recognition methods generally rely on extracting different motion representations of body parts to capture the gait information. However, the differences in appearance silhouettes between subjects are not discriminative enough to identify the distinct gaits. Therefore, exploring the representations of unique motion differences of subjects from the temporal domain is an effective way to boost gait recognition performance. In this paper, we propose a cross-view spatiotemporal aggregation network (CSTAN) to learn the most discriminative motion representations of subjects from multi-scale features. Particularly, CSTAN uses temporal contextual feature aggregation (TCFA) block to interact multi-scale features, increasing the diversity of temporal representations in feature sequences and producing suitable motion representations. Besides, a potential spatial feature assessment (PSFA) block is proposed to address the problem of missing silhouettes due to occlusion. Considering the multi-view context, we propose a cross-view prediction learning (CVPL) block to mitigate the impact of the view itself. By aggregating features in the spatiotemporal and view domains, CSTAN can learn a more appropriate representation of human motion differences. Extensive experiments on two public gait recognition datasets demonstrate the state-of-the-art performance of our proposed CSTAN. It is worth mentioning that the performance of CSTAN achieves a breakthrough in some complex environments, e.g., the subjects wearing different coats or carrying a bag.