Pedestrian Trajectory Extraction Method Based on UAV Video

Abstract

With the development of UAV technology, pedestrian trajectory extraction based on UAV video plays an increasingly prominent role in public safety. Aiming at the problems of small pedestrian targets in UAV video and the effect of pedestrian detection and tracking is easily blocked by obstacles in the scene, this paper analyzes the multi-target tracking algorithm framework based on detection, takes YOLOv5 as the target detection model, and gets a better anchor frame through KMeans++ clustering method. At the same time, CBAM attention module is integrated into YOLOv5 network to realize the effective extraction of small target features; With Deep_SORT is a target tracking model. By calculating the confidence of the trajectory, an improved correlation matching tracking algorithm framework is proposed to solve the problems of trajectory fracture caused by pedestrians being blocked, and provide a strong guarantee for extracting complete and reliable pedestrian trajectory information in UAV video. The mAP of this paper is 49.1 % on the VisDrone2019-DET dataset, and the MOTA is 48.0% on the VisDrone2019-MOT dataset. Experiments show that the pedestrian trajectory extraction method in this paper can extract more stable and continuous pedestrian trajectory.