Adaptive graph structure learning for skeleton-based gait recognition

Gait recognition, which identifies individuals based on their walking patterns, offers a contactless and robust biometric approach. Skeleton-based gait recognition using graph convolutional networks (GCNs) is favored for its robustness against external variations, such as clothing and carried objects. However, existing methods typically rely on static graph structures and lack thorough investigation into graph partitioning, failing to capture implicit, fine-grained inter-joint dependencies. Since human walking involves complex coordination beyond natural anatomical links, adaptively learning graph topologies and partition patterns is crucial for uncovering discriminative features. Therefore, we propose AGSL-Gait, an Adaptive Graph Structure Learning framework that learns data-driven residual topology refinement to capture fine-grained inter-joint dependencies. Specifically, we introduce an adaptive adjacency matrix that learns structural relationships directly from data. In addition, we explore multiple partition strategies to better model complex body connections and optimize the graph topology. Furthermore, we construct MoCap-Gait, a multi-view motion capture dataset tailored for forensic identification scenarios. Experimental results demonstrate that the proposed method achieves competitive performance across three distinct datasets: the widely used CASIA-B, the large-scale OUMVLP-Pose, and the proposed MoCap-Gait dataset.