Optimization Design of 3D Posture Reconstruction of Multi-view Dance Videos Under the Background of Artificial Intelligence

Under the background of rapid development of artificial intelligence technology, 3D posture reconstruction of multi-view dance videos still faces problems such as occlusion, perspective difference and insufficient posture estimation accuracy. Therefore, this paper proposes an optimization design method to solve the above problems, aiming to improve the accuracy and stability of 3D posture reconstruction of dance videos. The methods include the following: (1) a multi-view synchronous acquisition system can be constructed to obtain comprehensive video data; (2) a convolutional neural network (CNN) can be used to detect 2D key points in multi-view videos; (3) 3D posture reconstruction can be performed by combining multi-view geometric constraints with deep learning; (4) a posture optimization module based on timing information can be introduced to reduce jitter and errors in the reconstruction process. The method achieved an average of 41.7 mm on MPJPE and 37.5 mm on P-MPJPE, showing excellent robustness and adaptability. The optimization design effectively solves the key problems in 3D posture reconstruction of multi-view dance videos, and provides new technical ideas and practical paths for posture analysis in complex dance scenes.