YOLOv8-based Dense Pedestrian Detection Algorithm
Zheng Cao,Yuzhou Yang,Zhuanzhuan Niu,Mingyue Zhao,Li Xu
DOI: https://doi.org/10.1109/ICCCS61882.2024.10603346
2024-04-19
Abstract:In scenarios with high pedestrian density, intricate backgrounds, and variable lighting conditions, pedestrian targets may overlap and appear small, leading to potential missed or false detections during the detection process. To address this, an enhanced YOLOv8-based pedestrian detection algorithm is proposed. Initially, YOLOv8 serves as the foundational model architecture. Building upon this, the RepVGGBlock module replaces the original Conv module, enhancing feature representation and facilitating accurate pedestrian identification in dense scenes. Furthermore, the Soft-NMS algorithm is employed in place of the traditional NMS algorithm, effectively reducing the missed detection rate of overlapped and occluded targets, thus improving model accuracy and recall rate. Additionally, CoordAttention is integrated into the YOLOv8 backbone network to mitigate overlap and occlusion issues by effectively integrating global and local information. Experimental results on the CrowdHuman dataset demonstrate precision of 82.92%, recall of 71.34%, mAP@.5 of 79.94%, and mAP@.5:.95 of 58.81%. Comparative analysis against the benchmark YOLOv8 model reveals improvements of 1.61% in precision, 6.78% in recall, 5.06% in mAP@.5, and 11.07% in mAP@.5:.95, indicating its enhanced suitability for dense pedestrian object detection tasks.
Computer Science