Multi-person interaction action recognition based on co-occurrence graph convolutional network

Abstract

Skeleton-based human action recognition has attracted more and more attention with the accessibility of depth sensors and the development of pose estimation technology. The recognition of multi-person interaction behavior with skeleton data has wide applications in real life. However, the existing skeleton-based action recognition methods mainly focus on single-person actions, and lack the research on the semantic information of the skeleton co-occurrence relationship in the spatial dimension. Here, a two-stream co-occurrence graph convolutional network (2s-CGCN) is proposed for multi-person interactive action recognition. The topological connection structure of the proposed model establishes connections between the human limbs and the head to capture the co-occurrence feature. In addition, the coordinates and co-occurrence information of the joints are simultaneously input into the model. The experiments of mutual actions in the NTU-RGB+D dataset show the effectiveness of the proposed model.