A Human Tracker Based on YOLOv5s and the Extended Kalman Filter
YOLOv5s's pre-training weight on the COCO dataset can detect human targets,while failing to recognize specific human po-ses. A human tracker based on YOLOv5s and the extended Kalman filter is proposed. When YOLOv5s can detect the human object,the extended Kalman filter is initialized. When YOLOv5s cannot detect,the extended Kalman filter is used to track,and the candidate re-gion is created in the prior results. The optimal candidate region is matched by using the different hash algorithm as the observation val-ue,to update the target position and realize the continuity of human detection. The experimental results show that in terms of tracking accuracy,the overlap rate between the proposed tracker and the"real boundary box"is 50.65％,and the center position error is 51.78 pixels. The real-time performance is better than that of the KCF and TLD trackers,and the frame rate is 26 frames per second faster than that of the KCF tracker.
万方数据
维普
