基于路侧激光雷达的目标识别方法研究  

Target Recognition Method Based on Roadside Lidar

作　　者：尹涛 傅雪晴 范俊德 刘晓庆 刘晓雷 田雨欣 Yin Tao;Fu Xueqing;Fan Junde;Liu Xiaoqing;Liu XiaoLei;Tian Yuxin(Jinan Xiantou Urban Development Investment Group Co.,Ltd.,Jinan 250000,China;School of Qilu Transportation,Shandong University,Jinan 250000,China;Construction Management Department of Jinan Start-up Area Management Committee,Jinan 250000,China;School of Civil Engineering,Tianjin Chengjian University,Tianjin 300000,China)

机构地区：[1]济南先投城市发展投资集团有限公司,山东济南250000 [2]山东大学齐鲁交通学院,山东济南250000 [3]济南新旧动能转换起步区管理委员会建设管理部,山东济南250000 [4]天津城建大学土木工程学院,天津300000

出　　处：《市政技术》2025年第3期93-100,共8页Journal of Municipal Technology

基　　金：国家重点研发计划(2022YFB2602102);泰山学者工程专项经费资助(11560082363123)。

摘　　要：为提升车路协同的环境感知能力,基于路侧激光雷达提出基于栅格索引的帧间差分背景滤除方法和一种基于高斯核函数的DBSCAN目标聚类算法。首先,利用直通滤波对点云进行过滤,提取道路区域点云作为感兴趣区域(ROI);其次,将ROI进行二维栅格化处理,建立栅格索引,进而利用帧间差分法计算相同索引的栅格密度差值,去除背景点云;最后,基于高斯核函数构建自适应系数函数,对DBSCAN算法近邻点聚类阈值的选取进行优化。在真实道路上提取了15000帧连续数据进行试验测试,结果显示:改进的DBSCAN聚类算法提高了点云簇的类内一致性和类间差异性,有效抑制了目标的过分割和丢失现象,车辆目标检测的平均精度为98.8%,平均召回率为93.6%。研究成果可为车路协同的环境感知提供技术支持。In order to improve the environmental sensing ability of vehicle-to-everything(V2X),an inter-frame differential background filtering method based on raster indexing and a DBSCAN target clustering algorithm based on Gaussian kernel function are proposed based on roadside Lidar are proposed.Firstly,the point cloud is filtered using straight pass filtering to extract the road area point cloud as the region of interest(ROI);Secondly,the ROI is 2D rasterized to establish the raster index.The raster density difference of the same index is calculated by the interframe difference method to remove the background point cloud;Finally,the adaptive coefficient function is constructed based on the Gaussian kernel function to optimize the selection of the nearest-neighbor point clustering threshold of the DBSCAN algorithm.15000 frames of continuous data are extracted on roads for experimental testing.The results show that the improved DBSCAN clustering algorithm improves the intra-class consistency and interclass variability of point cloud clusters,effectively suppresses the over-segmentation and loss phenomenon of targets,with the average precision of the vehicle target detection of 98.8%,and the average recall rate of 93.6%.The research results can provide technical support for the environment sensing of vehicle-road collaboration.

关 键 词：激光雷达 背景滤除 点云聚类 DBSCAN算法 车路协同 

分 类 号：TN958.98[电子电信—信号与信息处理] TP391.4[电子电信—信息与通信工程]