Research on human skeleton behavior recognition method based on GCN-Transformer hybrid network architecture

In response to the problem of insufficient global information extraction from human skeletons during the spatial domain modeling process based on Graph Convolutional Networks (GCN) for human skeleton behavior recognition , we have proposed a hybrid network architecture that integrates GCN and Transformer for human skeletal behavior recognition. Firstly, a hybrid network structure is designed, which combines the advantages of GCN in local feature extraction and Transformer in global information capture, enabling comprehensive and effective extraction of global and local spatial features from skeletal data. Secondly, a spatial feature interaction module is constructed to dynamically adjust attention information in space and time, better fusing global and local features. Finally, we constructed an end-to-end graph convolutional neural network model that significantly improved the stability and accuracy of behavior recognition by enhancing the information representation ability of spatio-temporal receptive fields. Verified on the public MSR Action 3D action recognition datasets, the recognition rate of the proposed algorithm reaches 96.75%, surpassing most existing action recognition technologies.