SAGCN: Self-Attention Graph Convolutional Network for Human Pose Embedding

Accurate human pose feature representation is essential for action recognition. While traditional convolutional neural networks (CNNs) have achieved significant progress in extracting pose features, they are limited in capturing structural relationships and long-range dependencies between distant keypoints, and exhibit reduced robustness under keypoint occlusion. To overcome these limitations, we propose SAGCN, a novel model integrating graph convolutional network (GCN) with the self-attention mechanism. SAGCN effectively preserves the structural relationships among keypoints through GCN and captures long-range dependencies using self-attention. Furthermore, we introduce probabilistic embedding to represent the uncertainty inherent in multi-view human poses. The proposed SAGCN was evaluated on cross-view pose retrieval tasks using the Human 3.6M and MPI-INF-3DHP datasets, and its video sequence alignment capabilitywas assessed on the PenAction dataset. Experimental results showthat SAGCN not only outperforms existing methods in cross-view pose retrieval but also achieves competitive results in sequence alignment compared to specialized approaches.