Robust Detection and Tracking Algorithm for Pedestrian Targets with Posture Change

Abstract

AbstractIn the intelligent monitoring system, pedestrian detection and tracking is the basis of behavioral analysis, pedestrian retrieval and other intelligent analysis technologies. In order to solve the problems caused by posture change and occlusion, this paper proposes a robust pedestrian target detection and tracking algorithm, which is based on YOLOv3 and DeepSORT. In order to reduce the impact of pedestrian posture change, we introduce a person re-identification feature that resists posture change on the tracking algorithm DeepSORT. For pedestrian occlusion, inspired by Spindle Net, we integrate seven local appearance features extracted from different regions of the human body together to form robust global appearance features, making the tracking more accurate. Experimental results prove that the algorithm proposed in this paper shows robust pedestrian detection and tracking performance on both subjective visual effects and objective metrics, which can meet the needs of actual applications.KeywordsSurveillance videoAnti-posture changePedestrian detectionPedestrian tracking