Pedestrian Target Tracking Based On DeepSORT With YOLOv5
Yuqiao Gai, Weiyang He, Zilong Zhou
Abstract
Pedestrian target tracking is an important problem in the field of computer vision. To address low tracking accuracy and tracking errors in pedestrian target tracking. This paper developed a YOLOv5-based DeepSORT pedestrian target tracking algorithm (YOLOv5-DeepSORT), which introduces the high-performing YOLOv5 algorithm into the DeepSORT algorithm, which detects the tracking video frame by frame, and then predicts the target position using Kalman filtering while matching the same target using the Hungarian algorithm. To verify the performance of the YOLOv5-DeepSORT algorithm, this paper compares it with the YOLOv3-DeepSORT algorithm for experiments. The results show that the YOLOv5-DeepSORT algorithm can detect and track pedestrians well. Compared with the YOLOv3-DeepSORT algorithm, the MOTA of the YOLOv5- DeepSORT algorithm is improved by 9.9%, the number of false detections and missed detections is significantly reduced, and the number of target ID switching is decreased by 23 times, which proves the effectiveness of the method in this paper.