A novel node and edge cyclic embedding graph convolutional network for skeleton-based two-person interaction...

Accurately recognizing two-person interactions is essential for social behavior analysis and public safety monitoring in daily life. Although human skeleton data has been widely used in two-person interaction recognition, how to effectively capture interactive information in dyadic actions remains a core challenge for the accurate recognition of two-person interactions. Given the natural advantage of Graph Convolutional Networks in processing human skeleton data, we propose a Node and Edge Cyclic Embedding Graph Convolutional Network for two-person interaction recognition. Existing GCN-based approaches for two-person interaction recognition usually treat node and edge features as mutually independent. The rich interactive information inherent in two-person interaction has been overlooked. To address the difficulty of effectively capturing interaction information, we propose a Tri-Graph Cyclic Block. Furthermore, in order to achieve effective temporal modeling of skeleton sequence data, we employ the widely used Multi-Scale Temporal Convolution Block to process sequential data by modeling temporal dynamics at multiple scales. Finally, to highlight the key features that characterize coordinated local body movements between the two interactors, we design a Tri-Graph Attention Block that takes multi-graph node features as guidance. The proposed method achieves an accuracy of 99.4±0.2% on the SBU-Kinect dataset, and achieves recognition accuracies of 95.2±0.2%, 97.6±0.2%, 90.7±0.3%, and 90.8±0.2% under different evaluation protocols on the interaction subsets of the NTU RGB+D 60 and NTU RGB+D 120 datasets, respectively. The proposed method obtains competitive recognition accuracy among the comparison methods. The code is available at https://github.com/JLiu920/NECEGCN.