Under the rapid development of artificial intelligence (AI) technology, this study constructs an AI-driven dance movement personalized training model based on the problems of poor movement recognition accuracy and insufficient personalized instruction in traditional dance training. Methodologically, a two-branch twin supervised learning model is used to realize 2D to 3D skeletal keypoint conversion, and the ST-GCN network is improved by incorporating spatio-temporal attention mechanism to enhance feature extraction in spatial and temporal dimensions. A dataset is constructed using 3,500 images extracted from concert and dance videos, containing six dance movement types, such as crossing the waist, lifting high, spreading one arm, waving, spreading both arms, and walking. The results show that the improved ST-GCN model achieves a recognition accuracy of 93.63% on the test set, which is 14 percentage points higher than the traditional residual network model, and the top-1 metric after fusing spatio-temporal attention is 86.66%, which is 5.63 percentage points higher than the original ST-GCN model. The conclusion shows that the proposed AI-driven dance movement recognition model can effectively solve the problems of movement occlusion and perspective change, significantly improve the recognition accuracy, and provide technical support for personalized dance training and health management.
