汽车工程 ›› 2021, Vol. 43 ›› Issue (7): 1005-1012.doi: 10.19562/j.chinasae.qcgc.2021.07.007

• • 上一篇    下一篇

基于自适应阈值的三维激光点云地面分割算法研究

张凯1,2,于春磊3,赵亚丽4,陈义飞2,杨蒙蒙3,江昆3()   

  1. 1.辽宁工业大学汽车与交通工程学院,锦州 121000
    2.北京超星未来科技有限公司,北京 100080
    3.清华大学车辆与运载学院,北京 100084
    4.清华大学电子工程系,北京 100084
  • 收稿日期:2020-11-16 修回日期:2021-01-27 出版日期:2021-07-25 发布日期:2021-07-20
  • 通讯作者: 江昆 E-mail:jiangkun@mail.tsinghua.edu.cn
  • 基金资助:
    国家重点研发计划(2018YFB0105000);国家自然科学基金联合基金(U1864203);青年基金(61903220);北京市科技计划课题(Z181100005918001);第66批中国博士后科学基金面上资助项目(2019M660622)

Research on Ground Segmentation Algorithm Based on Adaptive Thresholds for 3D Laser Point Clouds

Kai Zhang1,2,Chunlei Yu3,Yali Zhao4,Yifei Chen2,Mengmeng Yang3,Kun Jiang3()   

  1. 1.School of Automotive and Transportation Engineering,Liaoning University of Technology,Jinzhou 121000
    2.Novauto(Beijing)Co. ,Ltd. ,Beijing 100080
    3.School of Vehicle and Mobility,Tsinghua University,Beijing 100084
    4.Department of Electronic Engineering,Tsinghua University,Beijing 100084
  • Received:2020-11-16 Revised:2021-01-27 Online:2021-07-25 Published:2021-07-20
  • Contact: Kun Jiang E-mail:jiangkun@mail.tsinghua.edu.cn

摘要:

针对自动驾驶汽车感知模块中三维激光点云前后背景分割过程存在的误分割问题,提出了一种基于路面波动幅度的自适应阈值地面分割方法。该方法将原始点云进行栅格划分,依据点云数量信息设计了相应的高度阈值分割算法和地面平面模型分割算法。具体地,地面平面模型分割算法首先选取局部区域点集拟合地面平面模型,然后针对地面点云分割中存在的误分割问题,构建路面波动幅度方程,并采用基于点集分布特征的自适应阈值方法实现初步分割,最后借助分割后的地面点云重新优化平面模型与分割阈值。本文中基于开源语义分割数据集Semantic?KITTI提出了统一的算法评价基准数据集Semantic?Nova与性能评价指标,同时基于自研的自动驾驶汽车平台采集的实际场景进行了性能测试。试验结果表明,本文中提出的自适应阈值地面分割算法不仅在基准数据集上能够达到较高的精度,而且在实际场景中满足鲁棒性和实时性要求,具有较高的工程应用价值。

关键词: 三维激光点云, 自适应分割阈值, 地面分割, 平面模型

Abstract:

To address the problem of false background segmentation on 3D Lidar point clouds in the perception module of autonomous vehicles, an adaptive threshold segmentation method based on fluctuation range of road surface is proposed. At first, the original point cloud is divided into grids and the corresponding height threshold segmentation algorithm and the ground plane model segmentation algorithm are designed according to the number of points in a certain grid cell. Specifically, the ground plane model is fitted to a local point cloud subset inside the grid cell. Afterwards, an equation of road surface fluctuation is constructed for the problem of false segmentation in ground point cloud segmentation. Based on these point set distribution characteristics, an adaptive threshold method is used to realize an initial segmentation. Finally, the segmented point cloud subset is used to optimize the ground plane model and the segmentation threshold in an iterative manner. The paper proposes a unified benchmark dataset Semantic?Nova based on the open dataset Semantic?KITTI and performance evaluation indicators. Meanwhile, the performance test is conducted based on the actual scenes collected by the self?developed autopilot vehicle platform. The test results show that the adaptive threshold ground segmentation algorithm proposed in this paper can achieve high accuracy in benchmark dataset. Furthermore, it can meet the requirements of robustness and real?time applications in actual scenes, which has high engineering application value.

Key words: 3D laser point cloud, adaptive segmentation threshold, ground segmentation, plane model