OpenStreetMap-Based LiDAR Global Localization in Urban Environment Without a Prior LiDAR Map

Younghun Cho, Giseop Kim, Sangmin Lee, Jee-Hwan Ryu

TL;DR

This paper introduces a novel global localization method for autonomous vehicles that matches real-time LiDAR scans against publicly available OpenStreetMap (OSM) data, completely eliminating the need for pre-built LiDAR maps. By converting both OSM building footprints and LiDAR scans into 1D distance descriptors, the method achieves localization accuracy comparable to state-of-the-art deep learning techniques that rely on dense, computationally expensive prior maps.

What Problem Does This Paper Address?

For an autonomous vehicle to navigate, it must know exactly where it is. Traditionally, this is solved by driving a mapping vehicle through an area to collect massive 3D point clouds, building a dense "prior LiDAR map," and then matching real-time LiDAR scans against it.

However, building and maintaining these LiDAR maps is a massive logistical bottleneck. Every road must be pre-driven, and maps quickly become outdated as environments change. The authors tackle this by asking: Can we localize a vehicle globally (the "kidnapped robot" problem) using only free, globally accessible 2D maps like OpenStreetMap? While previous OSM-based localization attempts relied heavily on cameras or required complex particle filters to track movement over time, this paper aims for a standalone, single-frame LiDAR solution.

Methods

The core idea is to extract only the building information from both the map and the sensor, and compare them using a custom, lightweight descriptor.

  1. OSM Descriptor Generation: The system interpolates points along OSM roads. At each point, it casts virtual rays in 360 degrees and calculates the distance to the nearest OSM building footprint. This creates a 1D array (360 elements) representing the "skyline" of buildings from that exact spot.
  2. LiDAR Descriptor Generation: For a live LiDAR scan, the authors use a semantic segmentation network (RangeNet++) to isolate only the points belonging to buildings. They then calculate the shortest distance to these building points at every angle, creating a 1D array formatted identically to the OSM descriptor.
  3. Rotation-Invariant Keys: Because a vehicle might be facing any direction, directly comparing the 1D arrays is difficult. The authors transform the 1D descriptors into a 2D "context" (binning by distance and angle) and sum across the angles. This creates a "key" that describes the surrounding buildings regardless of which way the car is facing.
  4. Two-Stage Matching:
  5. Stage 1 (Fast Retrieval): The system compares the rotation-invariant LiDAR key against all OSM keys in the database to quickly find the top 200 candidate locations.
  6. Stage 2 (Fine Matching): For these top 200 candidates, the system takes the original 1D LiDAR descriptor, physically rotates it through all 360 degrees, and compares it to the 1D OSM descriptors to find the absolute best match and the vehicle's orientation.

Datasets

The authors evaluated their method using two standard autonomous driving benchmarks:

  • KITTI Odometry Dataset: Sequences 00, 02, 05, 06, 07, 08, 09, and 10.
  • KITTI-360 Dataset: Sequences 00, 05, 06, and 09.
  • OpenStreetMap (OSM): Standard map data corresponding to the geographic areas of the KITTI datasets, specifically utilizing the 'lines' (roads) and 'multi polygons' (buildings) layers.

Key Insights and Ideas

  • Buildings are the ultimate anchor: Dynamic objects (cars, pedestrians) and semi-static objects (trees, fences) ruin map matching. By aggressively filtering the LiDAR scan to only look at buildings, you can bridge the gap between a rich 3D sensor and a flat 2D map.
  • Decoupling rotation saves massive compute: Brute-force checking every map location at every possible vehicle rotation is too slow. By creating a rotation-invariant "key" for the first pass, the algorithm filters out 98% of the map instantly, allowing the computationally heavier rotational alignment to only run on highly probable candidates.
  • Deep learning isn't always necessary for global localization: While methods like PointNetVLAD use heavy neural networks to extract features from point clouds, this purely geometric, human-interpretable 1D descriptor performs just as well in urban environments.

Key Results

  • Accuracy: On KITTI sequence 00, the proposed method achieved a 48.34% Top-1 accuracy, slightly outperforming the state-of-the-art deep learning baseline PointNetVLAD (47.90%). This is a massive win considering PointNetVLAD requires a prior LiDAR map to function.
  • Intersection Robustness: The rotation-invariant design proved highly effective at intersections. In KITTI-360 sequence 00, the method achieved over 95% accuracy at intersections where the vehicle approached from entirely different directions than it had previously.
  • Speed: The two-stage pipeline is highly efficient. It takes about 0.23 seconds per frame on a standard Intel i7 CPU, whereas PointNetVLAD requires a high-end GPU to achieve similar speeds.

Limitations and Future Work

The authors are transparent about where the method breaks down:

  • Occlusions: If a building is blocked by a tall fence, dense trees, or a hill, the LiDAR hits the obstacle while the OSM descriptor expects a building. This mismatch causes localization failures.
  • Repetitive Environments: Long, straight roads with identical building setbacks create identical descriptors, confusing the algorithm.
  • Building-less Areas: The method fundamentally fails on highways or off-road environments where there are no buildings to reference.

Future Work: The authors suggest integrating dynamic point removal algorithms (like ERASOR) to better clean the LiDAR scans before descriptor generation, and exploring ways to make the system robust in environments where buildings are sparse or heavily occluded.

Summary updated

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.
BibTeX
@article{cho2022openstreetmap,
  author = {Younghun Cho and Giseop Kim and Sangmin Lee and Jee-Hwan Ryu},
  title = {{OpenStreetMap-Based LiDAR Global Localization in Urban Environment Without a Prior LiDAR Map}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2022},
  url = {https://openreview.net/forum?id=6DK7i94t5AG},
}

Loading related content...

Figure