Know Your Maps: Uncertainty-Aware Semantic LiDAR Localization with Dual Maps

This paper presents a novel GNSS-free semantic LiDAR localization algorithm for autonomous vehicles. It uses standardized HD and open-source maps to ensure localization both inside and outside the HD map. In the back-end, two localization strategies are applied: a GraphSLAM, which performs Simultaneous Localization and Mapping within a sparse HD map, and a Monte Carlo Localization (MCL) in an uncertainty aware, low-resolution semantic occupancy grid map, which is derived from OpenStreetMap. An efficient, general panoptic segmentation is the basis of the front-end of both localization strategies. While we subsequently extract geometric primitives for the GraphSLAM, the MCL uses downs amp led semantic point clouds. We introduce a coupling of GraphSLAM and MCL which aims to increase the robustness of the localization. In general, we show a robust and accurate localization in and outside the HD map. We evaluate our localization with real-world experiments, using our testing vehicle and an HD map of our campus. We reach a mean position error of 0.12 m inside the HD map and prevent the GraphSLAM from drifting.