Two-person interactive action recognition based on hypergraph convolutional networks

Abstract Two-person interactive action recognition has broad application prospects in security monitoring, educational assistance etc. Currently, the recognition methods based on the combination of joint point data and graph convolutional networks have made satisfactory progress. However, this type of method still has the problem that the behavioral feature of two-person interaction is not sufficient to represent action information, and the model is also very complex, which cannot be practically applied. To address these problems, a two-person interactive action recognition algorithm based on hypergraph convolutional networks is proposed. Firstly, the input data are symmetrically processed and enhanced, then the hypergraph structure is used to build the two-person interaction hypergraph, which includes the two-person hypergraph part and the two-person interaction relation matrices part. Finally, the two-person interaction hypergraph is deployed into a graph convolutional network with multiple stream inputs, and the Spatial-Temporal Part Attention(STPA) is added to realize the recognition of two-person interaction actions. With fewer parameters, the algorithm can capture both adjacent and long-distance spatiotemporal information within an individual, as well as better utilize interaction information between two people. Experiments on the NTU RGB+D interactive dataset show that the recognition accuracy of this algorithm is 98.24\%, and the inference speed of the model is faster than other methods.