Robust Localization for Intelligent Vehicles Based on Compressed Road Scene Map in Urban Environments

Abstract

High-precision localization is an important task for intelligent vehicles. Previous studies have used the raw point cloud for mapping and localization. The problems with this method are the large storage requirement of the map for localization in large-scale areas, and the many meaningless points that do not contribute to localization, and may even degenerate localization results, and result in drift and failure in the map matching process. This paper presents a robust localization method for intelligent vehicles based on the proposed compressed road scene map to solve these problems. This work consists of three parts: mapping, map matching, and sensor fusion. For mapping, the point cloud is first converted to 3D occupancy grids, and the grids are projected on a plane perpendicular to the ground and parallel to the longitudinal direction. This will produce a 2D grid map, which we name a compressed road scene map because it can reflect the depth of the road scene. For map matching, a Monte Carlo framework is used to sample several hypothetical real-time compressed road scenes, and the normalized information distance is used to measure the similarity between the hypothetical real-time compressed road scenes and the compressed road scene map. For sensor fusion, vehicle motion data provided by the IMU and wheel encoders are fused with the map matching result. The experimental results show that the storage of the compressed road scene map decreases substantially compared with the raw point cloud. The robustness and precision of the proposed localization method is also demonstrated via real-vehicle experiments in large-scale areas.