Analysis of action trajectory of armed police enemy capture based on computer vision and deep learning

Capturing the enemy is an important training subject of the Armed Police Force. The traditional action analysis method relies on manual observation, which is time-consuming and labor-intensive, and subjective. The purpose of this study is to improve the accuracy and automation of action analysis by using computer vision and deep learning technology, and to provide scientific basis for the training of armed police. Kinect V2 and inertial sensor (IMU) are used for data acquisition, and accurate synchronization of multiple devices is realized through synchronization trigger module. The acquisition environment is equipped with standardized training ground and multi-light source array, and the ground is equipped with grid coordinate system for calibration. Data preprocessing includes time stamp interpolation alignment, composite filtering noise reduction, spatial normalization and time normalization. In the aspect of feature extraction, the dynamic characteristics of joint angle and trajectory curvature are calculated, and the depth features are extracted by using the improved ST-GCN (space-time graph convolution network) framework. Motion trajectory analysis uses dynamic time warping (DTW) to calculate the similarity, LSTM-Attention network to identify the motion pattern, and One-Class SVM to detect anomalies. Experimental results show that the average accuracy of this method is 92.3%, the trajectory alignment error is 4.2cm, and the end-to-end processing delay is 250 milliseconds, which is superior to the traditional HMM method and the benchmark 3D-CNN model. Confusion matrix analysis shows that the recognition accuracy of each action is over 88%, and the model is excellent in reducing confusion and improving recognition accuracy. Under extreme lighting conditions, the accuracy of the model drops to 85.6%, and the data diversity needs to be further increased. Real-time test shows that desktop GPU has fast processing speed but high power consumption, and edge devices are suitable for energy-sensitive scenarios.