A Geodetic Normal Distribution Map for Long-Term LiDAR Localization on Earth

Light detection and ranging (LiDAR) sensors enable a vehicle to estimate a pose by matching their measurements with a point cloud (PCD) map. However, the PCD map structure, widely used in robot fields, has some problems to be applied for mass production in automotive fields. First, the PCD map is too big to store all map data at in-vehicle units or download the map data from a wireless network according to the vehicle location. Second, the PCD map, represented by a single origin in the Cartesian coordinates, causes coordinate conversion errors due to an inaccurate plane-orb projection, when the vehicle estimate the geodetic pose on Earth. To solve two problems, this paper presents a geodetic normal distribution (GND) map structure. The GND map structure supports a geodetic quad-tree tiling system with multiple origins to minimize the coordinate conversion errors. The map data managed by the GND map structure are compressed by using Cartesian probabilistic distributions of points as map features. The truncation errors by heterogeneous coordinates between the geodetic tiling system and Cartesian distributions are compensated by the Cartesian voxelization rule. In order to match the LiDAR measurements with the GND map structure, the paper proposes map-matching approaches based on Monte-Carlo and optimization. The paper performed some experiments to evaluate the map size compression and the long-term localization on Earth: comparison with the PCD map structure, localization in various continents, and long-term localization.