Dynamic segmentation algorithm of dance video image for non-legacy ethnic dance inheritance

Traditional manual arrangement struggles to efficiently handle complex backgrounds, frequent movements and variable lighting in dance videos; high-precision automatic techniques are urgently needed for digital analysis and protection. In this paper, a large-scale dance video dataset covering multi-ethnic, multi-scene and multi-illumination conditions is constructed, and the preliminary extraction of foreground region is realised by using the motion detection module combining frame difference method and optical flow method. Then, based on the improved U-Net structure, multi-scale feature fusion and attention mechanism are designed to enhance the segmentation ability of clothing and limb details, and the joint loss of Dice and cross entropy is used to improve the boundary accuracy. Experimental results show that the proposed method is better than U-Net and DeepLabv3+ in terms of IoU, Dice, precision, recall, F1-score, etc., and shows stronger robustness and near real-time processing speed in complex scenes.