一种基于细粒度车辆检测与跟踪的公路交通调查方法  

A highway traffic survey method based on fine grained vehicle detection and tracking

作　　者：张森 侯效伟 詹春兰 李宁 单君 齐冲冲 ZHANG Sen;HOU Xiaowei;ZHAN Chunlan;LI Ning;SHAN Jun;QI Chongchong(Yunnan Provincial Highway Science and Technology Research Institute,Kunming 650051,China;School of Information Science and Engineering,Yunnan University,Kunming 650504,China;Yunnan United Vision Technology Co.,Ltd.,Kunming 650504,China)

机构地区：[1]云南省公路科学技术研究院,云南昆明650051 [2]云南大学信息学院,云南昆明650504 [3]云南联合视觉科技有限公司,云南昆明650504

出　　处：《应用科技》2024年第5期114-121,共8页Applied Science and Technology

基　　金：云南省交通运输厅项目(2022-23-3)。

摘　　要：在大多数国省道路场景下,基于计算机视觉的交调方式都面临着车型划分精细、道路交通参与者类型繁杂、噪声干扰多等,导致车型流量占比计算困难的问题。为此,本文提出了一种基于细粒度目标检测与跟踪的九型车识别框架,引入基于无锚框和密集特征采样的实时目标检测器(real-time models for object detection,RTMDet)作为检测模块来执行高效、精准的九型车检测任务;同时设计了一种具有任务针对性的感兴趣区域(region of interest,ROI)噪声抑制模块,用于过滤背景噪声和路面无效车辆。通过进一步与深度简单在线和实时跟踪(deep simple online and realtime tracking,DeepSort)框架集成,本文在检测和跟踪精度方面相较于主流方法都得到了提升,可以为二级交调任务提供精准、细粒度的道路流量信息。Within the context of most national and provincial road scenarios,computer vision-based traffic survey encounters difficulties in precise vehicle categorization,intricate classification of road traffic participants,and susceptibility to noise interference,consequently posing challenges in accurate vehicle flow ratio estimation.In light of these challenges,this paper introduces a novel framework for the recognition of nine distinct vehicle categories,employing fine-grained object detection and tracking.It incorporates RTMDet,a detection module based on anchor-free principles and dense feature sampling,which facilitates high-efficiency and precise recognition of the nine vehicle categories.Additionally,a task-specific Region of Interest(ROI)noise suppression module has been devised to filter out background noises and irrelevant vehicles on the road surface.By further integrating with the DeepSort tracking framework,this study achieves enhancements in both detection and tracking accuracy compared to the mainstream methods.These advancements empower the provision of accurate,fine-grained road traffic information,rendering it valuable for secondary traffic monitoring tasks.

关 键 词：车辆检测 跟踪模型 交通调查 九型车检测 计算机视觉 交通量统计 实时目标检测器 深度简单在线和实时跟踪 

分 类 号：U495[交通运输工程—交通运输规划与管理]