AST-GCN: An Attention-Guided Spatial-Temporal GCN Framework for Gait Recognition

Gait recognition, as a biometric modality, faces significant challenges due to variations in covariates like clothing, carrying conditions, and viewing angles. Appearance-based approaches rely on silhouette images, but their vulnerability to outer aspects limits their robustness. Model-based approaches that utilize skeletal data demonstrate increased resilience. However, these approaches do not frequently capture the complex joint coordination. In addition, the traditional spatial-temporal graph convolutional networks (ST-GCN) exhibit limitations related to static spatial partitioning issues. This paper presents AST-GCN, a model-based gait recognition framework that effectively addresses existing limitations by integrating frame and channel attention (FCA) with ST-GCN. The FCA attention module coordinates relationships among physically disjoint joints and temporal frames, enhancing the discriminative power of gait features. The proposed method is assessed using the CASIA-B dataset, achieving rank-1 accuracies of 95.8%, 91.1%, and 87.3% under normal walking (NM), carrying bag (BG), and wearing coat (CL) conditions, respectively, indicating superior performance. AST-GCN demonstrates notable advancements relative to current methodologies, especially in challenging covariate conditions, underscoring its resilience to variations in clothing and viewing angles.