A Graph Skeleton Transformer Network for Action Recognition

Skeleton-based action recognition is a research hotspot in the field of computer vision. Currently, mainstream method based on Graph Convolutional Networks (GCNs). Although there are many advantages GCNs, GCNs mainly rely graph topologies to draw dependencies between joints, which limited capturing long-distance dependencies. Meanwhile, Transformer-based methods have been applied skeleton-based because they effectively capture However, existing lose inherent connection information human skeleton joints do not yet focus initial structure information. This paper aims improve accuracy recognition. Therefore, Skeleton Transformer network (GSTN) for proposed, architecture extract global features, while using undirected represented by symmetric matrix local features. Two encodings utilized feature processing joints’ semantic and centrality In process multi-stream fusion strategies, grid-search-based used assign weights each input stream optimize results. We tested our three datasets: NTU RGB+D 60, 120, NW-UCLA. The experimental results show that model’s comparable state-of-the-art approaches.