LiDAR-based HD Map Localization using Semantic Generalized ICP with Road Marking Detection

TL;DR

This work tackles GPS-denied autonomous vehicle localization by combining LiDAR-based road-marking detection with HD-map registration. It introduces an online pipeline that uses adaptive high-reflectance segmentation, a spatio-temporal probabilistic local map, and LiBEV images to semantically label road markings, followed by SG-ICP to register markings to the HD map with category-aware covariance, modeling linear markings as 2D 1-manifolds to reduce under-constrained effects. The approach achieves real-time performance (sub-50 ms average) and centimeter-level lateral localization accuracy across multiple LiDAR types and scenarios, with robustness to speeds and adverse weather. While effective, it does not handle roads without road markings and plans to leverage above-ground information to further enhance reliability in such cases.

Abstract

In GPS-denied scenarios, a robust environmental perception and localization system becomes crucial for autonomous driving. In this paper, a LiDAR-based online localization system is developed, incorporating road marking detection and registration on a high-definition (HD) map. Within our system, a road marking detection approach is proposed with real-time performance, in which an adaptive segmentation technique is first introduced to isolate high-reflectance points correlated with road markings, enhancing real-time efficiency. Then, a spatio-temporal probabilistic local map is formed by aggregating historical LiDAR scans, providing a dense point cloud. Finally, a LiDAR bird's-eye view (LiBEV) image is generated, and an instance segmentation network is applied to accurately label the road markings. For road marking registration, a semantic generalized iterative closest point (SG-ICP) algorithm is designed. Linear road markings are modeled as 1-manifolds embedded in 2D space, mitigating the influence of constraints along the linear direction, addressing the under-constrained problem and achieving a higher localization accuracy on HD maps than ICP. Extensive experiments are conducted in real-world scenarios, demonstrating the effectiveness and robustness of our system.

Figures (7)

• Figure 1: (a) The HD map localization of our approach is visualized, where the trajectory of vehicle localization is marked in green, and the current pose of the vehicle is represented by a red cube. The blue point cloud represents ground points from a single-frame LiDAR data. These ground points are adaptively segmented to identify highly reflective points. Subsequently, they are aggregated by successive frames of data to form a denser point cloud. Finally, semantic segmentation is applied to obtain a semantic point cloud, which is then registered with the HD map to estimate the vehicle's pose. (b) Road markings extracted using our approach are visualized, encompassing dashed lanes, solid lanes, stop lines, texts, arrows, diamond signs, triangle signs, curbs, and crosswalks.
• Figure 2: The flowchart of the proposed approach.
• Figure 3: The experimental scenarios (top) and their corresponding HD maps (bottom). (a) Fangshan1 (b) Jiashan (c) Fangshan2 (d) Airport.
• Figure 4: Comparison between the trajectories estimated by SG-ICP and ICP is conducted using ground-truth trajectories provided by RTK. The substantial localization error of SG-ICP and ICP are marked with purple and red lines, respectively, where estimated distance errors exceed 2.0 m or yaw errors surpass 5.0°.
• Figure 5: A box plot illustrating the time consumption for each sub-step of the proposed approach.