A target tracking system based on multi-camera information fusion

In order to promote the development of the public security sky-eyes system and the Internet of things, our system completes the integrated multi-camera cooperative tracking of multi-targets. The system uses the self-calibration method to map the camera detection results to the global coordinates through perspective transformation. The system uses YOLO-X network for target recognition and classification; The improved SORT algorithm is used to track and improve the ID switch problem. At the same time, 3D-box method is used to regress the attitude angle information of the object, so as to judge the state of the object and correct the deviation. On the premise of ensuring stability, the system achieves a faster running speed. It uses an RTX3080 GPU to accelerate, the overall speed reaches 60Hz, the classification accuracy is close to 100%, and stable target tracking is achieved. The system algorithm fully considers the occlusion, the change of target feature state, the disappearance of the target, etc., and can still ensure good results.