路侧多源感知数据集规范化构建方法  

Standardized constructing method of a roadside multi-source sensing dataset

作　　者：李立[1] 鲍宇健 杨文臣 楚庆玲 汪贵平[1] LI Li;BAO Yu-jian;YANG Wen-chen;CHU Qing-ling;WANG Gui-ping(School of Electronic and Control Engineering,Chang'an University,Xi'an 710064,China;National Engineering Laboratory For Surface Transportation Weather Impacts Prevention,Broadvision Engineering Consultants Co.,Ltd.,Kunming 650200,China)

机构地区：[1]长安大学电子与控制工程学院,西安710064 [2]云南省交通规划设计研究院有限公司陆地交通气象灾害防治技术国家工程实验室,昆明650200

出　　处：《吉林大学学报(工学版)》2025年第2期529-536,共8页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(71901040);陕西省自然科学基础研究计划项目(2023-JC-YB-507);云南交投科技研发项目(YCIC-YF-2022-06)。

摘　　要：为满足路侧多源融合感知算法研究对标准公开数据的需求,提出一种路侧多源感知数据集规范化构建方法。在城市T型交叉口采集激光雷达和图像数据并进行时空匹配,提出包括道路空间划分、路面分割和激光点云聚类等步骤的车辆三维外形尺寸提取方法,提出涵盖目标过滤和分类、识别难度划分、三维边界框校准、标签信息补充等步骤的车辆标注方法,构建了昼夜条件下含有9794个小汽车和重车标签的规范化路侧多源感知数据集;使用YOLOv5算法和PointRCNN算法对本文数据集的车辆二维和三维目标识别效果进行测试。测试结果表明:由于场景复杂度、采集设备以及车辆类型的差别,本文数据集与公开车载数据集中车辆平均激光点数量、车辆三维边界框尺寸方面存在明显差异;YOLOv5算法和PointRCNN算法对路侧多源感知数据集中的车辆目标具有与公开车载数据集相近的目标识别精度。To meet the need of standard open datasets for the researches of roadside multi-source fusionsensing algorithms,this paper proposed a method to construct a standard roadside multi-source sensingdataset.The LIDAR and image data was collected at an urban T-junction and matched each other in bothspatial and temporal dimensions.A vehicle three-dimension configuration extraction method was proposed,which included the steps of road space division,road pavement segmentation,and laser point cloudclustering.A vehicle labeling method was designed,which included the steps of target filtering andclassification,recognition difficulty division,3D bounding box calibration,and tag information supplement.It constructed a standardized roadside multi-source sensing dataset that contained the labels of 9794 carsand heavy vehicles in daylight and nighttime.The YOLOv5 algorithm and the PointRCNN algorithm wereused to test the 2D and 3D target recognition performance on the constructed dataset.Test results showedthat due to the differences of scene complexity,data collection device and vehicle type,the constructeddataset and open vehicular datasets had differences in terms of average number of scene and vehicle laserpoints,and size of vehicle 3D bounding box.The YOLOv5 algorithm and the PointRCNN algorithm havesimilar vehicle target recognition accuracy on the open vehicular datasets and the constructed roadsidemulti-source sensing dataset.

关 键 词：智能交通 路侧感知 数据集 多源感知 激光雷达 目标识别 

分 类 号：U495[交通运输工程—交通运输规划与管理]