Traditional dance teaching mainly relies on empirical transmission and lacks scientific means of movement analysis, making it difficult to accurately quantify human movement characteristics. In this study, we constructed a digital modeling system for human movement mechanism in dance anatomy based on 3DResNet-LSTM cascade neural network. Methodologically, Kinect somatosensory technology was used to collect dance movement data, extract the skeletal information of 24 joint points, and process the data noise through the skeletal joint point motion smoothing algorithm. A 28-layer 3DResNet-LSTM cascade network model is constructed, in which 3DResNets are responsible for extracting spatial features and the LSTM network learns temporal features to realize accurate recognition of multi-view dance movements. The results show that the model performance is optimized when the number of LSTM neurons is 24, the learning rate is 0.0055, and the batch size is 200. In the test of more than 100 dance action segments, the recognition accuracy reaches 98.66%, which is 7.83% higher than the average of STGCNs and 22.94% higher than the average of LSTM. In the application validation of four professional dance types, the recognition accuracies are above 89.53%, up to 93.08%, and the shortest recognition time is only 22.87 s. The 3DResNet-LSTM cascade network model in this study demonstrates excellent generalization ability and robustness in the task of dance movement recognition, which provides an effective technical support for the digital teaching of dance anatomy.
