Intelligent Choreography Algorithm Recorded by Optical Motion Capture System through Directed Graph Neural Networks

Dance and music are inextricably connected. At present, there are several problems in intelligent music-dance choreography, such as stiff movements, poor beats-alignment between dance and music, and difficulty in long sequences generation. Learning to model music-dance generation is yet challenging. To solve these problems, we propose a new intelligent music-dance choreography framework based on a Directed Graph Nueral Network, through an improved temporal directed graph and spatial directed graph, in which music features and motion features are regarded as control signals, respectively. Utilizing prior knowledge, we added the rotation restrictions of human skeleton to prevent the abnormal joints drift, and feet-end contact restrictions to fix foot slidings. Besides, we constructed a music-dance dataset containing two types recorded by an optical motion capture system. Experimental results demonstrate that the proposed method is workable and can generate higher quality dance sequences.