Abstract: Aiming at the problem of frequent identity exchange in multi-target tracking due to target occlusion, we proposed a pedestrian multi-target tracking algorithm. Firtsly, the algorithm used YOLOv4 as the detector to detect the target and determine the coordinates of the detection frame, the extended Kalman filter was used to predict the trajectory. Secondly, the Hungarian algorithm was used as the data association module, the detection frame predicted by extended Kalman filter was matched with the detection frame of target detection by cascade matching method, and the trajectory anomaly correction algorithm was added for the occluded target. Finally, the experiments were carried out on the test set of the MOT16 data set. The experimental results show that the algorithm achieves 56.5% tracking accuracy, and has a good effect on the occlusion phenomenon, which effectively improves the problem of frequent identity switching and target loss caused by occlusion.

Key words: computer vision, multi-target tracking, YOLOv4, extended Kalman filter