In order to quickly, accurately and stably recognize and track multiple targets in motion, this work builds a real-time video perception tracking system for multiple moving targets. Firstly, based on the original YOLOv4 target detection algorithm, the theoretical innovation of the backbone feature extraction network is carried out, which aims to improve the expression ability of the network and enhance the feature extraction ability, and secondly, the YOLOv4-PPMCC-aware tracking model is proposed. Given the original frame of the video, the tracking target template is selected, the improved YOLOv4 detection algorithm is used to sense multiple moving targets in the real-time video in real-time, and all the targets in the video frame are found out by paying attention to their position information. Then the similarity between the template and the target is calculated by Pearson correlation analysis (PPMCC), and the tracking object is the highest similarity with the template. Finally, the proposed model is deployed to the device for system experimental verification.