Abstract: To address the problems of complex action features,incomplete feature extraction and low recognition accuracy of existing action recognition methods in figure skating,a multi-scale spatio-temporal graph convolutional network sharing multibranch features and attention is proposed for action recognition in figure skating.First,the OpenPose algorithm is employed to extract skeletal point data of human motion to eliminate noise interference; second,the attention structure of the channel is improved,and the improved attention mechanism enables the model to extract more comprehensive and critical features; then,the multi-scale spatio-temporal graph convolution network with fused attention mechanism is constructed to extract more complete temporal features; finally,the shared features are extracted and fused into the network to allow the model to share data while mining semantic features.The results show that the recognition accuracy of the FSD-10 dataset with 30 types of figure skating movements is 64.5%.Compared with both ST-GCN and CTR-GCN methods,the algorithm achieves higher accuracy in all cases,indicating its effectiveness in the action recognition in figure skating.