Research on action recognition and intelligent optimization of armed police enemy capture based on machine learning

Capturing the enemy is very important to the actual combat ability of armed police officers and soldiers. Traditional training depends on the coach's experience, which has some problems such as low efficiency and difficult to quantify the effect. In this study, machine learning technology, especially deep learning algorithm, is used to collect multi-modal data such as training videos through multi-view RGB-D cameras. After data enhancement and other preprocessing, a feature engineering scheme is constructed by combining spatio-temporal feature extraction and attention mechanism. Double-stream spatio-temporal graph convolution network (GCN) model is designed to process action data, and reinforcement learning (RL) module is introduced to optimize action performance, and a real-time feedback system is created to simulate and optimize the training process. The experiment was carried out in a 6m×6m training ground equipped with NVIDIA A100 GPU and other hardware, and 8 kinds of core action data of 120 soldiers were collected. The results show that the accuracy of this model is 93.4%, the key frame alignment error is 19.8mm, and the delay is 68ms. After optimization, the kinetic energy efficiency of action is improved, and the risk of joint injury is reduced. Compared with the traditional training mode, the intelligent optimization system significantly shortens the training time, improves the action standard and the winning rate of actual combat simulation.