Optimization of Wushu Sanshou Technical Movement Recognition and Training Based on Machine Learning Algorithm

In order to further improve the recognition rate and optimize the efficiency of wushu sparring action judging, this paper designs a wushu sparring action recognition model based on extracting the advantages of the skeletal point action recognition model with ResNet50 as the basic framework, adding the self-attention mechanism and CBAM attention module. Faster-RCNN is applied as a human body detector to label the human body's region range, and the continuous attention (CA) structure is used as a residual link structure to optimize the recognition accuracy. Construct the Wushu Sanshou dataset and test the action recognition model with dual attention mechanism. Combine the essentials of Wushu Sanshou movements with the skeletal point characteristics of human posture to propose core muscle group stability training for Sanshou movements. Two groups (experimental group and control group) were trained for ten weeks to compare the dynamic and static performance of Wushu Sanshou movements before and after the training. After the core muscle stability training, the performance of the experimental group was significantly different from that of the pre-training group in terms of dynamic (time to reach stabilization in the vertical/forward/backward direction) and static (total offset of the center of pressure, maximum offset in the forward/backward/left/right direction), which illustrated the effectiveness of the core muscle stability training of the movement based on the essentials of the Wushu sparring movement.