Attention Interactive Graph Convolutional Network for Skeleton-Based Human Interaction Recognition

Feng Gao, Hailun Xia, Zhihao Tang

Abstract: Skeleton-based Human interaction action recognition is a great challenge in the field of human activity analysis. The existing graph convolutional networks (GCNs) based recognition methods ignore the spatial-temporal interactive features between two persons. Therefore, we propose Attention Inter-active Graph Convolutional Network(AIGCN) to explore hu-man interaction. First, we design an Interactive Attention En-coding GCN(IAE-GCN) module to extract interactive spatial structures, reflecting their influence on each other by regarding joint position encoding as semantic information. Second, we propose Interactive-Attention Mask TCN (IAM-TCN) that extracts temporal interactive features, representing temporal attentional excitation signal among the different joints of dif-ferent people along the time. Finally, we verify the effectiveness of our method on classic human interaction datasets SBU and interactive action sub-datasets of NTU-RGB+D and NTU-RGB+D 120, and experiment shows that our models achieve state-of-the-art performance.