高速公路车道级视频检测区自动设定方法  

An Automatic Configuration Method for Video-Based Lane-Level Detection Zones on Expressways

作　　者：莫宇蓉 吴烈阳 彭锦辉 刘圣卿 唐先亮 黎川 符锌砂[3] MO Yurong;WU Lieyang;PENG Jinhui;LIU Shengqing;TANG Xianliang;LI Chuan;FU Xinsha(Traffic Monitoring&Command Center of Jiangxi Province,Nanchang 330046,China;BYD Auto Industry Co.,Ltd.,Shenzhen 518118,China;School of Civil Engineering and Transportation,South China University of Technology,Guangzhou 510641,China)

机构地区：[1]江西省交通监控指挥中心,江西南昌330046 [2]比亚迪汽车工业有限公司,广东深圳518118 [3]华南理工大学土木与交通学院,广东广州510641

出　　处：《交通运输研究》2024年第5期78-90,共13页Transport Research

基　　金：国家自然科学基金项目(51978283);江西省交通运输厅科技项目(2022X0037)。

摘　　要：针对云台摄像枪在轮巡过程中无法自动设定视频检测区,从而影响交通事件识别准确率的问题,提出了一种基于视频的高速公路车道级检测区自动设定方法。首先,通过分析U-Net和MobileNet系列模型的特征,结合深度可分离卷积和倒置残差等结构,设计了一个高效且轻量化的R-Net系列模型,专门用于车道线和可行驶区域的语义分割。在此基础上,根据高速公路特定场景识别任务的特点,提出了一种基于连通域分析的车道线和检测区标记算法,实现了车道级检测区的自动设定。同时,为了提高标记算法的准确率,首次引入了阈值处理和叠加帧数这两种预处理方法,然后利用二次方程对车道线标记结果进行拟合,实现了车道线完整且平滑的分割。实验结果表明,R-Net系列模型的性能指标MIoU与传统模型如SegNet和U-Net接近,但显著减少了模型参数量和内积运算量,其中R-NetV2模型的分割性能指标MIoU达到90.6%,与U-Net相比仅下降了0.4%,但其模型参数量减少了38.7%,内积运算量减少了62.5%。对经过预处理后的语义分割结果进行标记,车道线标记准确率与传统方法相比从80.47%提高到95.58%。The paper proposed an automatic configuration method for lane-level detection zones on expressways based on video,addressing the issue of the inability of Pan-Tilt-Zoom cameras to automatically set up video detection zones,which affected the accuracy of traffic event recognition.Firstly,by analyzing the features of U-Net and MobileNet series models,combined with structures such as depthwise separable convolution and inverted residuals,a lightweight and efficient R-Net series model was designed specifically for semantic segmentation of lane lines and drivable areas.On this basis,according to the characteristics of expressways'specific scene recognition tasks,an algorithm for lane lines and detection zones labeling based on connected component analysis was proposed,which achieved automatic configuration of lane-level detection zones.At the same time,in order to improve the accuracy of the labeling algorithm,two preprocessing methods,threshold processing and frame stacking,were introduced for the first time.Then,a quadratic equation was used to fit the lane labeling results,achieving complete and smooth segmentation of the lane lines.The experimental results showed that the performance metric MIoU of the R-Net series models was close to traditional models such as SegNet and U-Net,but significantly reduced the number of model parameters and inner product operations.Among them,the segmentation performance metric MIoU of the R-NetV2 model reached 90.6%,which was only 0.4%lower than U-Net,but its model parameter count was reduced by 38.7%and inner product operation was reduced by 62.5%.Labeling the preprocessed semantic segmentation results resulted in an increase in lane labeling accuracy from 80.47%to 95.58%compared to traditional methods.

关 键 词：交通事件识别 视频检测 车道级检测区 轻量化 语义分割 连通域 车道线 

分 类 号：U495[交通运输工程—交通运输规划与管理]