面向结构化场景的激光雷达点云高精度配准与定位方法

High Precision Registration and Positioning Method of LIDAR Point Cloud for Structured Scene

关键词:  激光雷达  平面模型  特征直方图  线性匹配

Key words:  Lidar  Plane model  Feature histograms  Linear match

摘要:

在基于先验地图的激光雷达室内导航方案中，通常采用点云配准的方法进行无人设备位姿初始化。在结构化场景下，传统配准算法特征鲁棒性较差，导致点云配准的误差较大且易陷入局部最优。针对该问题，提出了一种基于多平面空间模型的点云快速配准方法。首先该方法利用特征直方图的思想对空间点云进行快速粗聚类，根据平面一致性将粗聚类后的点集进行合并形成面特征，从而对密闭空间进行平面模型化表示。随后通过空间平面排序实现了面特征的快速关联，并利用线性匹配方法实现了两帧点云的精确配准，从而解算出机体在先验地图中的相对位姿。最后通过Gazebo搭建的仿真环境与室内结构化模拟环境对算法进行了验证。结果表明，在大型结构化场景下，算法具有更好的适应性以及更高的计算效率，能够快速为无人系统提供精准的地图初始位姿。

Abstract:

In the lidar indoor navigation scheme based on a prior map, the point cloud registration method is usually used to initialize the position and pose of unmanned equipment. In the structured scenario, the poor robustness of the traditional algorithm leads to a large error in point cloud registration and a tendency to local optimality. To solve this problem, a fast point cloud registration method based on multi-plane space model is proposed. Firstly, this method uses the idea of feature histograms to conduct rapid rough clustering of spatial point clouds. According to the consistency of the plane, the point sets after rough clustering are combined to form the surface features, so as to carry out the plane model representation of the confined space. Then the spatial plane model is used to realize the rapid correlation of surface features, and the linear matching method is used to realize the accurate registration of two frame point clouds, so as to solve the relative pose of the body in the prior map. Finally, the algorithm is verified by the simulation environment built by Gazebo and the indoor structured simulation environment. The results show that in large structured scenarios, the algorithm has better adaptability and higher computing efficiency, and can quickly provide accurate initial pose for unmanned systems.