LIDAR MATCHING STRATEGIES FOR HD POINT CLOUD MAP GENERATION IN URBAN AREA

Abstract

Abstract. Autonomous driving relies on high accuracy point vector map, which was generated by the point cloud map, and pre-provides the vehicle preliminary road environment information. Lidar Odometry and Mapping (LOAM) has always been a promising research topic in the field of robotics, environment sensing, and currently autonomous driving. However, in certain urban environments like basement parking lot, tunnels, highways, or other similar settings, the geometric features are not clearly discernible. As a result, algorithms resembling the LOAM framework may encounter difficulties in accurately mapping these areas. The paper utilized relative low-cost LiDAR expecting to propose a state-of-the-art point cloud mapping/update scheme. We compared the GNSS-challenge area with straight line and loop area separately, simultaneously considered the DG, ICP, NDT matching algorithm for the low-cost mapping/update strategy. With the realistic experiment conduction, our result evaluated by point to point corresponding mean error and standard error. For the straight line environment, ICP has the fastest convergence in empirical cumulative distribution under 0.4 meters. For the loop scenario, point-to-point ICP still has the fastest convergence in empirical cumulative distribution under 0.22 meters. Yet both of them still suffer from the fault matching.