In this paper, a VR virtual training system integrating judo sleeping technique and police combat is designed to realize immersive combat simulation through multimodal VR platform. The proposed algorithm is based on the weighted joint position difference to determine the positive and negative sides of the skeleton, combined with the dynamic adjustment mechanism of joint confidence and the bone vector synthesis method to restore the human skeleton in the training state. In the visual motion capture module, an improved CMPs iterative architecture is adopted to enhance the data feature extraction capability through convolutional kernel replacement, residual dense connection and two-branch PAF-confidence map co-training. It is shown that the joint training of multiple feature subsets leads to a model recognition accuracy of 84.27%. The power law normalization process drives the accuracy to exceed 90%, and the lowest error rate is only 10.605% in 60 times of motion capture recognition. Utilizing platform-assisted training enabled the police officer’s technical level and comprehensive physical fitness score to exceed 90 points.
