Study on Multi-Pedestrian Trajectory Tracking based on improved YOLOv5+DeepSort

Abstract

A multi-object tracking algorithm based on improved YOLOv5+Deepsort is proposed to improve the tracking effect in crowded and fuzzy scenes. The algorithm is improved as follows: the SKNet visual attention mechanism is integrated into the Backbone of YOLOv5 to strengthen the ability of recognizing fuzzy crowded groups; the FPN+PAN structure of the feature fusion module of YOLOv5 is replaced with the BiFPN structure to achieve efficient bidirectional cross-scale connectivity and weighted feature fusion;finally,constantly velocity model in the Kalman Filter is replaced with the constantly acceleration model to optimize the pedestrian motion model,and use DIOU quadratic matching to match detection frames that are not matched successfully,to improve the DeepSort tracking performance. The experimental results show that the accuracy on MOT17 is improved by 5.20% and the precision is improved by 1.85%; the accuracy on MOT20 is improved by 4.09% and the precision is improved by 1.33%.