基于改进YOLOv5算法的道路交通参与者实时检测方法  

A Method of Real-time Detection for Road Traffic Participants Based on an Improved YOLOv5 Algorithm

作　　者：张逸凡 聂琳真 黄灏然 尹智帅 ZHANG Yifan;NIE Linzhen;HUANG Haoran;YIN Zhishuai(College of Automotive Engineering,Wuhan University of Technology,Wuhan 430070,China;Hubei Engineering and Technology Center of New Energy and Intelligent Connected Vehicle,Wuhan University of Technology,Wuhan 430070,China;Dongfeng Commercial Vehicle Technical Center,Dongfeng Commercial Vehicle Co.,Ltd.,Wuhan 430056,China)

机构地区：[1]武汉理工大学汽车工程学院,武汉430070 [2]新能源与智能网联汽车湖北省工程技术中心,武汉430070 [3]东风商用车有限公司东风商用车技术中心,武汉430056

出　　处：《交通信息与安全》2024年第1期115-123,共9页Journal of Transport Information and Safety

基　　金：湖北省重点研发计划项目(2022BAA081)资助。

摘　　要：从道路监控图像中快速准确地检测交通参与者对于智能交通系统监管道路目标具有重要意义。为解决传统YOLOv5目标检测算法对多种交通参与者目标检测精度低、重叠目标漏检等问题,研究了基于改进YOLOv5算法的道路交通参与者实时检测方法。为增强浅层网络提取图像特征信息能力,采用融合移动翻转瓶颈卷积(FusedMBC)代替原卷积结构,并通过自注意力机制学习交通参与者的纹理特征;为加强主干网络感知图像空间特征信息的能力,引入坐标注意力机制(CA),使主干网络更加关注图像中交通参与者的语义特征;为使普通卷积拥有感知构造能力,以增强激活空间的灵敏度,采用漏斗激活函数(FReLU)作为卷积层的激活函数,并能够使特征向量进行像素级建模;为增强网络对密集目标的空间特征信息提取能力,在特征融合网络中加入坐标注意力机制,通过注意力捕捉密集目标融合后的空间与通道特征信息,让网络精确定位各个目标。通过对车路协同自动驾驶数据集DAIR-V2X的交通参与者图像进行数据增强预处理,构建用于验证模型性能的测试集2000张并进行了算法验证。实验结果表明:①改进后的YOLOv5算法平均检测精度达到82.4%,平均召回率达到95%,平均检测速度达到204帧/s。②相比于原始YOLOv5,其在平均检测精度和平均检测速度分别提高了5.8%和33.3%,证实提出的方法能够实现快速准确地检测交通参与者,有助于提升智能交通系统监管交通参与者的能力。Rapidly and accurately detecting traffic participants from road surveillance images is of great significance for intelligent transportation systems to monitor road targets.With the aim of solving the issues low detection accuracy and disability of detecting overlapping targets of the original YOLOv5 algorithm for various traffic participants,a real-time detection method of road traffic participants based on an improved YOLOv5 algorithm is proposed.To improve the capacity of shallow network to extract image characteristics,the fused mobile inverted bottleneck convolution(FusedMBC)is adopted to replace the original convolution structure to speed up the reasoning speed of the shallow neural network,and the self-attention mechanism is used to learn the texture features of traffic participants To enhance the ability of backbone network to perceive spatial features of images,the coordinate attention mechanism(CA)is introduced,which makes the backbone network pay more attention to the semantic characteristics of traffic participants in the images.To enable conventional convolution to capture visual layouts and enhance the sensitivity of activation space,the funnel activation function(FReLU)is adopted as the activation function of the convolution layer,and the feature vector can be modeled at the pixel level.To enhance the ability of extracting spatial features for dense targets,a coordinate attention mechanism is introduced to the feature fusion network,which captures the spatial and channel feature information of densely fused targets through attention mechanism,the network can accurately locate each target.Through data enhancement preprocessing on images of traffic participants based on the data set DAIR-V2X about vehicle-road cooperative and autonomous driving,a test set of 2000 images is developed to verify the property of the model.Experimental results show that:①The improved YOLOv5 algorithm has a mean average precision of 82.4%,an average recall rate of 93%,and an average detection speed of 204 frames/s.②In c

关 键 词：智能交通 交通目标 交通参与者检测 YOLOv5 融合移动翻转瓶颈卷积 坐标注意力机制 

分 类 号：U495[交通运输工程—交通运输规划与管理]