Abstract: In view of the low application of key joints and features in human action recognition based on skeleton data, an improved action recognition system based on the fusion of spatial-temporal graph convolution neural network and channel-spatial union attention block is proposed. The structural features were obtained by spatial graph convolution, and the key joints and key structure information were enhanced by channel-spatial union attention module. The advanced spatial-temporal features were obtained by time graph convolution. The recognition results were obtained by global pooling layer and Softmax classifier. The experimental results show that while the key joints and structural features are enhanced, the original feature information is retained. This algorithm has higher accuracy in skeleton-based action recognition.