Overlapped Pedestrian Detection Based on YOLOv5 in Crowded Scenes

Abstract

Pedestrian detection in a crowded environment is challenging for vehicle intelligent driving systems. At present, pedestrian detection algorithms have achieved great performance in detecting well-separated figures. However, pedestrians are generally overlapped in crowded scenes, resulting in slow detection speed, low detection accuracy, and poor robustness in pedestrian detection technology. In this paper, the YOLOv5 algorithm is used for pedestrian detection. In the aspect of data pretreatment, Mosaic data enhancement, unified image size, adaptive anchor frame calculation, and other processing are carried out for data.YOLOv5 can detect targets at multiple scales, and CIOU_Loss and DIOU_nms are applied to the YOLOv5 algorithm. It can improve the recognition ability of the occlusion target and has a good detection effect on the detection of the occlusion pedestrian target through the training network of amplified data set. The verification experiment shows that the pedestrian detection model based on YOLOv5 has great detection accuracy and recall rate in detecting covered pedestrians.