Application of Video Swin Transformer in Modern Dance Choreography Style Generation and Innovation Assistance

Given the complex movements, lack of fixed structure, and highly variable emotional expressions of modern dance, existing models struggle to effectively identify and capture diverse stylistic features. Similar spatiotemporal representation challenges are also encountered in intelligent electromagnetic information processing and signal understanding tasks. This paper proposes a modern dance style modeling method based on the Video Swin Transformer. First, modern dance video data are collected, and skeletal information is extracted to achieve strict spatiotemporal synchronization and spatial alignment between RGB video frames and corresponding human skeletal keypoint sequences. Then, hierarchical spatiotemporal dance features are extracted using the Video Swin Transformer architecture, and a style attention mechanism is designed to construct interpretable style encoding vectors. Subsequently, visual dimensionality reduction is applied to the low-dimensional style space to analyze the distribution of dancer movement styles and establish style mappings. Finally, combined with a generative decoder and a style control interface, the style vectors are used for dance movement reconstruction and creative generation, completing a full cycle from analysis to innovation assistance. Experimental results demonstrate that the proposed model achieves an average classification accuracy of 80.5% across different styles, while the optimal Euclidean distance for style space separation reaches 1.42 with an overlap rate of only 3.8%, validating its effectiveness in style control and generation and providing useful insights for spatiotemporal representation learning in intelligent signal processing.