基于车载激光点云的道路几何信息自动化提取

Automated Extraction of Road Geometry Information Using Mobile LiDAR Point Cloud

为了高效地进行道路设施信息采集与数字化建模,利用车载激光点云数据构建了一种自动化提取道路几何信息的方法框架。针对激光数据的无序性和冗余性,通过网格降采样和半径滤波精简点云规模、去除噪音点;通过栅格单元划分进行点云组织和索引,合理利用点云的空间局部性、缩减运算规模;利用道路要素在高程上的层次性与路面结构的连续性、光滑性,设计了高程滤波、基于主成分分析框架的局部法向量滤波、DBSCAN聚类等方法,实现从原始点云到路面点云的精确分割;利用采集车辆的行驶轨迹信息获取道路走向,利用其方向向量与法向量进行道路横截面的划分;切取横截面后投影至二维平面,并通过滑动窗口、最小二乘等算法提取道路宽度与平纵横参数。通过提取算法与人工测量的结果对比,在复杂街区和郊区公路两个实验数据集,点云分割准确性均超过87%,完整性均超过97%,提取质量均超过86%,几何信息的平均相对误差较小,说明算法具有良好的提取质量。有限算力条件下,两个数据集中点云处理时间分别是6.864与10.078 s/km,几何信息提取时间分别是1.732和0.843 s/km。提出的方法能够很好的兼顾提取效率与精度,在复杂街区和郊区公路环境下具有良好的适用性,可为道路设施的健康评定和三维重建提供参考。

To efficiently collect and digitally model road facility information, this paper constructed a method framework for automatic extraction of road geometry information by using vehicle-mounted laser point cloud data.For the disorder and redundancy of laser data, grid drop sampling and radius filtering were used to simplify the size of the point cloud and remove noise points. The point cloud is organized and indexed by grid cell division, and the spatial locality of the point cloud is rationally utilized to reduce the scale of operation. Using the hierarchy of road elements on elevation and the continuity and smoothness of pavement structure, elevation filtering, local normal vector filtering based on the principal component analysis framework, and DBSCAN clustering methods were designed to achieve accurate segmentation from the original point cloud to the pavement point cloud. The road direction was obtained by collecting vehicle trajectory information, and the road cross section was divided by its direction vector and normal vector. The cross-section was cut and projected onto a two-dimensional plane, and the road width and horizontal and horizontal parameters were extracted by sliding window and least square algorithm. By comparing the extraction algorithm with the manual measurement results, in the two experimental data sets of complex blocks and suburban roads, the accuracy of point cloud segmentation is more than 87%, the integrity is more than 97%,and the extraction quality is more than 86%. The average relative error of geometric information is small, indicating that the algorithm has good extraction quality. Under the condition of finite computation, the processing time of two data centralized point clouds is 6. 864 and 10. 078 s/km, respectively, and the extraction time of geometric information is 1. 732 and 0. 843 s/km, respectively. The proposed method can give a good balance between extraction efficiency and accuracy, and has good applicability in complex blocks and suburban highway environments. It can provide a reference for the health assessment and three-dimensional reconstruction of road facilities.