Skeleton action recognition based on spatio-temporal features

Graph convolutional networks have performed well in action recognition tasks using skeleton data, but topology extraction remains a challenge. Most methods fail to consider the correlations between channels in skeleton topology graphs. To address this, we propose a multi-channel optimal graph convolutional network (MOGCN) that combines multi-channel optimal graph convolution and multi-scale temporal convolution techniques. MOGCN exhibits improved capability in processing spatio-temporal information and node relationships, generating optimal skeleton topology graphs that model correlations between all nodes in the sequence. We also use multi-scale temporal convolution to extract temporal features, improving the modeling of long-term correlations and global correlations of spatio-temporal joints. We further improve accuracy using a multi-stream fusion network model. Experiments on multiple datasets show that the proposed MOGCN module is effective and improves the classification accuracy of skeleton-based action recognition.