基于YOLOv5s优化模型的道路交通标志异常状态检测  

Traffic Sign Abnormality Detection Based on Improved YOLOv5s Model

作　　者：陈子昂 郭唐仪[1] 隋立岩 周洋[1] 陈新[1] CHEN Zi-ang;GUO Tang-yi;SUI Li-yan;ZHOU Yang;CHEN Xin(School of Automation,Nanjing University of Science and Technology,Nanjing,Jiangsu 210094,China)

机构地区：[1]南京理工大学自动化学院,江苏南京210094

出　　处：《公路交通科技》2024年第10期8-16,共9页Journal of Highway and Transportation Research and Development

基　　金：国家重点研发计划项目(2019YFE0123800);南京市国际合作项目(202002013)。

摘　　要：道路交通标志的遮挡、磨损等异常,对交通安全和通行效率造成很大的影响。针对目前道路交通标志异常状态检测实践中存在的检测误差高、模型大而难嵌入设备终端等问题,提出了一种基于YOLOv5s优化的目标检测轻量化模型。提出的YOLOv5s优化模型内容包括:以MobileNetv3网络来替换模型主干网络,减小模型的大小;以BiFPN网络替换原生的PANet网络,减少冗余计算,提高模型特征融合的能力;以CBAM轻量级注意力机制代替MobileNetv3的SENet注意力机制模块,提升模型准确性;引入焦点损失函数,改善正负样本数量失衡影响。以南京、淮安获取的交通标志异常状态图片为基础数据,运用图像处理方法将原始的数据集进行扩充,共获取到了2 511张标志遮挡异常状态的图片和2 615张标志磨损异常状态的图片。运用该模型检测道路交通标志遮挡和磨损等异常,采用Adam优化器训练至模型收敛。结果表明,模型的mAP达到90.5%,精确度为91.28%,召回率为90.32%,F1分数为0.9,检测速度达到52帧/s。与原YOLOv5s模型相比,模型大小仅为其1/2,且精确度、F1分数、检测速度分别提高了3.84%,0.03,3.84帧/s。结果表明,提出的基于YOLOv5s优化的目标检测模型,在提高检测准确率的同时还能够满足轻量化的需求,能较好嵌入智能养护移动端,实现对于道路交通标志异常状态的智能化检测。Aiming at the problems of high detection error and too large model to embed equipment terminal in the current practice of traffic signs abnormality detection,the lightweight target detection model based on improved YOLOv5s was proposed.The proposed improved YOLOv5s model included replacing the model backbone network with MobileNetv3 network,and reducing the model size;replacing the native PANet with BiFPN to reduce the redundant computing,and improving the model feature fusion ability;replacing SENet attention mechanism module with CBAM lightweight attention mechanism to improve model accuracy;introducing the focus loss function to improve the influence of positive and negative sample size imbalances.Taking the abnormal pictures of traffic signs obtained in Nanjing and Huai’an as the basic data,and using the image processing method to expand the original data sets,the total of 2511 images of sign obscuring abnormal and 2615 images of sign wearing abnormal state were obtained.The model was used to detect anomalies,e.g.,road traffic sign occlusion and wear;and the Adam Optimizer was used to train the model to converge.The result indicates that the model mAP reaches 90.5%with the accuracy of 91.28%,the recall rate of 90.32%,the F 1 score of 0.9,and the detection speed reaching 52 frames per second.Compared with the original YOLOv5s model,the improved model size is only 1/2;and the accuracy,F 1 score,and detection speed increase by 3.84%,0.03,and 3.84 frames per second respectively.The proposed target detection model based on improved YOLOv5s can improve the detection accuracy while also meeting the needs of lightweight.It can better embed intelligent maintenance mobile terminals to achieve intelligent abnormality detection for traffic signs.

关 键 词：智能交通 异常检测 YOLOv5s 交通标志 轻量化 

分 类 号：U491.5[交通运输工程—交通运输规划与管理]