Research on Vision-based pedestrian detection and tracking algorithm

Abstract

Target detection and target tracking play an important role in many scenes. For example, intelligent monitoring, multi-target tracking in complex environments. This paper designs a pedestrian detection and tracking system. Based on Yolov5, the proposed detection algorithm introduces the deformable convolution to construct the Res-dcn component to replace the residual component of the Yolov5 backbone network, which can accurately locate the target feature points. Pedestrian tracking is based on DeepSort multi-target tracking algorithm, for the mismatch problem of pedestrian IDS in front and back frames in complex environment, the fusion of FHOG feature and CNN feature is introduced to the DeepSort multi-target tracking algorithm, the candidate feature matching mechanism is used to filter the trajectory set to improve the tracking performance. Experiment results with comparisons show the validity and the superiority of our developed tracking algorithm.