Spatial-Temporal Interleaved Attention Graph Convolutional Networks for Skeleton-Based Hand Action Recognition

In recent years, skeleton-based hand movement recognition has garnered significant attention with the advancement of graph convolutional networks. While previous approaches have concentrated on modeling joint dependencies and enhancing representation through various attention mechanisms, they often fall short in capturing information over long distances and extended time periods. To address this limitation, we introduce the STIA-GCN model. This model employs three parallel graph convolution blocks to extract diverse action patterns, followed by multi-scale temporal convolution to enrich feature representation. Additionally, spatial-temporal interleaved attention mechanism is incorporated to effectively capture long-range dependencies across different temporal frames and spatial locations, thereby facilitating a deeper understanding of complex action patterns. Evaluations on benchmark datasets SHREC'17 Track and DHG-14/28 reveal that the proposed method outperforms existing approaches in dynamic gesture recognition tasks.