[Paper Review] OpenStreetMap-based LiDAR Global Localization in Urban Environment without a Prior LiDAR Map

Younghun Cho, Giseop Kim+2 more|arXiv (Cornell University)|Feb 15, 2022

Remote Sensing and LiDAR Applications4 citations

TL;DR

This paper proposes a novel LiDAR-based global localization method that uses OpenStreetMap (OSM) building data to generate 1D descriptors for vehicle localization without requiring prior LiDAR maps. By comparing OSM-derived descriptors with real-time LiDAR scan descriptors, the method achieves comparable accuracy to deep learning-based LiDAR map methods like PointNetVLAD, outperforming it in top-1 accuracy on KITTI sequence 00.

ABSTRACT

Using publicly accessible maps, we propose a novel vehicle localization method that can be applied without using prior light detection and ranging (LiDAR) maps. Our method generates OSM descriptors by calculating the distances to buildings from a location in OpenStreetMap at a regular angle, and LiDAR descriptors by calculating the shortest distances to building points from the current location at a regular angle. Comparing the OSM descriptors and LiDAR descriptors yields a highly accurate vehicle localization result. Compared to methods that use prior LiDAR maps, our method presents two main advantages: (1) vehicle localization is not limited to only places with previously acquired LiDAR maps, and (2) our method is comparable to LiDAR map-based methods, and especially outperforms the other methods with respect to the top one candidate at KITTI dataset sequence 00.

Motivation & Objective

• To enable LiDAR-based global vehicle localization without relying on pre-built LiDAR maps, which are costly and hard to maintain.
• To address the limitation of existing OSM-based localization methods that require additional filtering or particle filters for robustness.
• To develop a descriptor-based method that is both efficient and accurate, using only static OSM building data and real-time LiDAR scans.
• To demonstrate that OSM-based descriptors can achieve performance comparable to deep learning-based methods using LiDAR maps.
• To validate the method on standard benchmarks like KITTI and KITTI-360, showing robustness and scalability in urban environments.

Proposed method

• Generate OSM descriptors by computing the shortest distance from each point on a road to the nearest building in OpenStreetMap at regular angular intervals (e.g., 1° or 2°).
• Extract LiDAR descriptors from real-time 3D point clouds by measuring the minimum distance to building points at the same angular resolution.
• Use a 1D descriptor representation that encodes the spatial layout of surrounding buildings, enabling rotation-invariant matching.
• Perform global localization by comparing the LiDAR descriptor from a query scan with all OSM descriptors in the database using a similarity metric.
• Apply a two-stage search: first, retrieve top 200 candidates based on descriptor similarity; second, refine the top result using a more precise matching step.
• Leverage the static nature of OSM building data to create a persistent, map-like reference without the need for LiDAR data collection.

Experimental results

Research questions

• RQ1 Can OSM-based descriptors alone enable accurate global LiDAR localization without prior LiDAR maps?
• RQ2 How does the performance of OSM-based localization compare to state-of-the-art deep learning methods like PointNetVLAD that rely on LiDAR maps?
• RQ3 In what scenarios does the OSM descriptor fail, and what environmental factors degrade its performance?
• RQ4 Can the proposed method achieve high top-1 accuracy without using particle filters or additional post-processing?
• RQ5 How efficient is the descriptor generation and matching process in real-time urban localization tasks?

Key findings

• The proposed method achieves a top-1 accuracy of 48.34% on KITTI dataset sequence 00, slightly outperforming PointNetVLAD’s 47.90%.
• The method shows comparable performance to PointNetVLAD in top-5 (56.86% vs. 61.73%) and top-10 (61.15% vs. 66.79%) accuracy, indicating strong global localization capability.
• The 1-stage method achieves 19.60% top-1 accuracy, which improves to 48.34% in the 2-stage method, demonstrating the effectiveness of candidate refinement.
• Descriptor generation time is 0.1714s per frame, comparable to PointNetVLAD’s 0.0690s (on GPU), despite running on CPU, indicating high efficiency.
• The method fails in environments with no buildings (e.g., highways) or where buildings are occluded by fences, trees, or hills, as LiDAR scans detect no structures.
• Repeated building patterns over long stretches reduce descriptor discrimination, leading to multiple similar candidates and localization ambiguity.