BDF-YOLOV5: Improved YOLOV5 Based on Bi-directional Fusion Network for Dense Pedestrian Detection

Pedestrian detection is a challenging task in the field of computer vision and plays a crucial role in downstream tasks, such as video surveillance and autonomous driving.Despite significant progress over the past two decades, scale variance and occlusion remain prominent issues. To address these problems, we propose BDF-YOLOv5 in this paper. Based on YOLOv5, we replace the original FPN with the BDF network structure. Furthermore, to further improve our BDF-YOLOv5, we additionally improved the loss function for bounding box regression and proposed weighted-CIOU. Extensive experimental results on the Crowdhuman dataset demonstrate the feasibility of our method. Compared to the baseline model (YOLOv5), BDF-YOLOv5 achieves an improvement of approximately 4.0%.