基于改进YOLOv5和DeepSort的交通流参数检测方法  

Traffic flow parameter detection method based on improved YOLOv5 and DeepSort

作　　者：单振宇 张琳 侯晓雯 Shan Zhenyu;Zhang Lin;Hou Xiaowen(School of Traffic and Transportation Engineering,Changsha University of Science&Technology,Changsha 410114,China;School of Information and Engineering,Ocean University of China,Qingdao 266100,China)

机构地区：[1]长沙理工大学交通运输工程学院,长沙410114 [2]中国海洋大学信息与工程学部,青岛266100

出　　处：《现代计算机》2023年第14期1-7,13,共8页Modern Computer

基　　金：长沙理工大学研究生科研创新项目(CX2021SS113)。

摘　　要：为缓解基于视频的交通流参数检测效果差的问题,提出了基于改进YOLOv5s和DeepSort的实时交通流参数检测方法。首先,在YOLOv5s中引入MobilenetV3和SIoU Loss来轻量化检测模型,提高检测速度。其次,重构Deep⁃Sort外观特征提取网络,并基于车辆重识别数据集重训练,提高跟踪准确率。最后,基于虚拟检测线圈的方式检测双向交通流量和车速。实验结果表明:方法整体运行速度可达59帧/s,在平峰、高峰时段的检测准确率均在92%以上,在实际交通流参数检测任务中具有较强的鲁棒性。Real‑time traffic flow parameter detection method based on improved YOLOv5s and DeepSort is proposed to alleviate the problem of poor video traffic flow parameter detection.First,MobilenetV3 and SIoU Loss are introduced in YOLOv5s to lighten the detection model and improve the detection speed.Secondly,the DeepSort appearance feature extraction network is lightened and retrained on the vehicle re‑identification dataset to improve the tracking accuracy.Finally,the two‑way traffic flow and vehicle speed are detected based on the virtual detection coil.The experimental results show that the overall operation speed of the algorithm can reach 59 fps,and the accuracy in both flat and peak is above 92%,which has strong robustness in the actual traffic flow parameter detection tasks.

关 键 词：交通流参数检测 YOLOv5s DeepSort 实时 

分 类 号：U491.11[交通运输工程—交通运输规划与管理]