面向车流量智能检测的YOLOv7算法改进与应用  

Improvement and Application of YOLOv7 Algorithm for Intelligent Detection of Traffic Flow

作　　者：马庆禄[1] 吴跃川 张梓轩 李杨梅 MA Qing-lu;WU Yue-chuan;ZHANG Zi-xuan;LI Yang-mei(School of Transportation,Chongqing Jiaotong University,Chongqing 4ooo74,China;Ningxia Jiaotou Expressway Management Co.Ltd.,Ningxia 750000,China)

机构地区：[1]重庆交通大学交通运输学院,重庆市400074 [2]宁夏交投高速公路管理有限公司,银川市750000

出　　处：《公路》2024年第1期242-249,共8页Highway

基　　金：宁夏回族自治区交通运输厅科技项目,项目编号NJGF20200301。

摘　　要：针对当前机器视觉识别中车流量检测的精度问题,以YOLOv7人工智能算法为基础,通过视觉跟踪并叠加注意力机制,提出一种基于YOLOv7和Deep SORT的改进型车流量智能检测方法。通过将注意力模块GAM与YOLOv7网络进行融合增强检测网络的特征提取能力;同时在改进后的YOLOv7网络中引入Deep SORT跟踪算法以改善车辆间相互遮挡导致复检漏检问题。实验选取重庆市渝中区经纬大道双向六车道为研究对象,在新铺社天桥上采用固定相机连接移动笔记本电脑的方式进行数据采集以及算法验证,为了保证算法的可重复性,分别选取早高峰、午平峰和晚高峰3个时段分别录取了5 min的交通流视频。利用在交通视频中通过设置虚拟检测线,让新算法在车辆检测的同时对车辆运行轨迹进行跟踪,当车辆经过检测线时记录车辆的身份编号,以此来实现交通视频的车流量监测与跟踪计数。实验结果表明:改进后的新算法相比于原YOLOv7算法在车辆检测方面平均精度提高了2.3%,视频车流量统计的精度提高了8.2%。Aiming at the accuracy issue of vehicle flow detection in current machine vision recognition,a modified vehicle flow intelligent detection method based on YOLOv7 artificial intelligence algorithm and Deep SORT is proposed by adding visual tracking and attention mechanism.The attention module GAM is fused with the YOLOv7 network to enhance the feature extraction capability of the detection network.Meanwhile,Deep SORT tracking algorithm is introduced into the improved YOLOv7 network to improve the problem of re-checking and missed detection caused by vehicle occlusion.In the experiment,the dualdirectional 3-lanes Jingwei Avenue in Yuzhong District of Chongqing is selected as the research object,and data collection and algorithm verification are conducted by connecting a fixed camera to a mobile notebook computer on the Xinpushe Overpass.To ensure the repeatability of the algorithm,5-minute traffic flow videos are recorded during the morning rush hour,noon off-peak period,and evening rush hour.By setting virtual detection lines in the traffic video,the new algorithm tracks the trajectory of vehicles while detecting them,and records the identity number of the vehicles when they passe the detection line to achieve traffic video flow monitoring and tracking counting.The experimental results show that compared with the original YOLOv7 algorithm,the modified algorithm has increased the average precision of vehicle detection by 2.3%and the accuracy of video vehicle flow statistics by 8.2%.

关 键 词：智能交通 车流量检测 YOLOv7 Deep SORT 深度学习 

分 类 号：U491.116[交通运输工程—交通运输规划与管理] TP183[交通运输工程—道路与铁道工程]