汽车工程 ›› 2022, Vol. 44 ›› Issue (7): 987-996.doi: 10.19562/j.chinasae.qcgc.2022.07.005

所属专题: 智能网联汽车技术专题-感知&HMI&测评2022年

• • 上一篇    下一篇

基于自适应阈值DBSCAN的路侧点云分割算法

金立生1,2,贺阳1,王欢欢1,霍震1,谢宪毅1,郭柏苍1()   

  1. 1.燕山大学车辆与能源学院,秦皇岛  066004
    2.燕山大学,河北省特种运载装备重点实验室,秦皇岛  066004
  • 收稿日期:2021-12-30 修回日期:2022-02-13 出版日期:2022-07-25 发布日期:2022-07-20
  • 通讯作者: 郭柏苍 E-mail:guobaicang@ysu.edu.cn
  • 基金资助:
    国家重点研发计划(2021YFB3202200)、国家自然科学基金(52072333)和河北省省级科技计划(21340801D,F2021203107)资助。

Point Cloud Segmentation Algorithm Based on Adaptive Threshold DBSCAN for Roadside LiDAR

Lisheng Jin1,2,Yang He1,Huanhuan Wang1,Zhen Huo1,Xianyi Xie1,Baicang Guo1()   

  1. 1.School of Vehicle and Energy,Yanshan University,Qinhuangdao  066004
    2.Yanshan University,Hebei Key Laboratory of Special Delivery Equipment,Qinhuangdao  066004
  • Received:2021-12-30 Revised:2022-02-13 Online:2022-07-25 Published:2022-07-20
  • Contact: Baicang Guo E-mail:guobaicang@ysu.edu.cn

摘要:

针对路侧采集的激光雷达点云数据随距离增大而密度下降导致同一目标的点云被分割成多个目标的问题,提出了一种基于自适应阈值DBSCAN的路侧点云分割算法。首先,使用改进GPF和直通滤波对采集的路侧点云进行过滤,提取出道路区域上的非地面点云;然后,基于有效距离和sigmoid函数构建自适应系数函数,对DBSCAN聚类算法集群生长中近邻点搜索时半径阈值的选取规则进行优化;最后,利用自适应阈值DBSCAN聚类算法对非地面点进行聚类,得到隶属于单个目标的点云。采集了1 055帧真实场景的连续数据进行测试,结果显示:C-H系数平均约增加3倍、D-B系数平均减少4.52%、轮廓系数平均增加77.78%,这表明基于自适应阈值DBSCAN的分割算法能提高点云簇的类内一致性和类间差异性,有效减少路侧激光雷达点云的过分割现象,具有较高的工程应用价值。

关键词: 智能交通, 点云聚类, DBSCAN, 自适应阈值, 目标检测

Abstract:

Aiming at the defect of lidar point-cloud data collected from roadside, that the point-cloud in the same target is divided into multiple targets caused by the reduction of density due to the increase of distance, a roadside point-cloud segmentation algorithm based on DBSCAN with adaptive threshold is proposed. Firstly, the collected roadside point-cloud is filtered by using improved GPF and straight-through filter, and the non-ground point-cloud is extracted from road area. Then, the adaptive coefficient function is constructed based on the effective distance and sigmoid function, and the selection rule of radius threshold in vicinal point search during clan grow in DBSCAN clustering algorithm is optimized. Finally, the non-ground points are clustered using the DBSCAN clustering algorithm with adaptive threshold, with the point-cloud subordinating to single target obtained. The continuous data of 1 055 frames of real scenes are collected for testing, and the results show that the average C-H coefficient increases by around three times, the average D-B coefficient rises by 4.52%, and the average contour coefficient is raised by 77.78%, indicating that the segmentation algorithm based on DBSCAN with adaptive threshold can enhance the intra-class consistency and inter-class difference of the point-cloud cluster, effectively reducing the over-segmentation phenomenon of roadside point-cloud with a high engineering application value.

Key words: intelligent transportation, point-cloud clustering, DBSCAN, adaptive threshold, target detection