基于改进YOLOv5的行人检测方法研究  

Research on pedestrian detection method based on improved YOLOv5

作　　者：薛继伟[1] 薛鹏杰 胡馨元 XUE Jiwei;XUE Pengjie;HU Xinyuan(School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China)

机构地区：[1]东北石油大学计算机与信息技术学院,黑龙江大庆163318

出　　处：《重庆理工大学学报（自然科学）》2024年第7期101-109,共9页Journal of Chongqing University of Technology：Natural Science

基　　金：黑龙江省省属本科高校基本科研业务费项目(2022TSTD-03)。

摘　　要：针对行人检测中出现的目标遮挡和小尺度目标漏检等现象,提出一种基于YOLOv5改进的行人检测模型DROE-YOLO。在YOLOv5的C3模块中引入了Res2Net的残差结构以增强网络对行人目标的表征能力。采用Dynamic Head作为YOLOv5的检测头,提高检测的准确性和鲁棒性。在标签分配策略方面采用了Simplified OTA方法,可以更准确地匹配真实框与预测框。最后,使用soft-NMS+EIOU的方法,进一步提高行人目标的检测准确率。在CrowdHuman数据集上的实验结果表明,DROE-YOLO在行人检测任务上取得了较好的效果。与基准模型相比,在增加少量参数的情况下,DROE-YOLO模型的检测精度提升了3.3%,召回率提升了6.5%,相比原模型更适用于实际的行人检测任务。To address target occlusion and missed detections of small-scale pedestrians in pedestrian detection,a modified pedestrian detection model called DROE-YOLO is proposed based on YOLOv5.Specifically,the residual structure of Res2Net is introduced into the C3 module of YOLOv5 to enhance the network’s representation capability for pedestrian targets.Additionally,Dynamic Head is employed as the detection head for YOLOv5 to improve detection accuracy and robustness.The Simplified OTA method is adopted for label assignment strategy,which enables more accurate matching between ground truth boxes and predicted boxes.Finally,the soft-NMS+EIOU method is used to further improve the detection accuracy of pedestrian targets.Our experimental results on the CrowdHuman dataset demonstrate that DROE-YOLO achieves excellent performances in pedestrian detection tasks.Compared to the baseline model,with a slight increase in parameters,DROE-YOLO model improves the precision by 3.3%and the recall by 6.5%,making it more suitable for practical pedestrian detection tasks.

关 键 词：行人检测 Res2Net Dynamic-Head Simplified-OTA Soft-NMS 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]