DP-YOLO: Enhancing Pedestrian Detection in Crowd Scenes with Deformable Convolution and Varifocal Loss
Li Jiao, Muhammad Irsyad Abdullah · 2024
Pedestrian detection in densely populated public places is of great significance. This study proposed a pedestrian detection algorithm named DP-YOLO to improve the detection performance. Focus on the problems such as small targets, occlusion, deformation, etc. DP-YOLO takes YOLOv5s as the baseline algorithm, replacing the original 4 downsampling standard convolutional layers with Deformable Convolution layers, using a dense object detection loss function Varifocal Loss, to improve the performance on occlusion and deformable targets, adding a extra Tiny Detection Head to capture more small targets. Experiments demonstrate, DP-YOLO is an effective pedestrian detection algorithm in crowded scenes, whose mAP50 achieves 0.897, 0.196 higher than the baseline algorithm YOLOv5s, far superior to the classic object detection algorithms and YOLO series lightweight classic algorithms.