GroundLoc: Efficient Large-Scale Outdoor LiDAR-Only Localization

AI summary

Overview

Research area: Mobile robotics, specifically LiDAR-based map localization for autonomous vehicles and robots in large-scale outdoor environments.

Technical level: Intermediate. The paper assumes familiarity with LiDAR sensors, bird's-eye-view (BEV) rasterization, keypoint matching, and trajectory error metrics, but the pipeline itself is described at a level accessible to graduate students in robotics.

One-sentence scope: The paper presents GroundLoc, an open-source pipeline that localizes a robot in large outdoor environments using only LiDAR, by matching learned or handcrafted keypoints between a 3-channel BEV image of the current ground surface and a compact 2D raster prior map.

What This Paper Is About

Mobile robots and autonomous vehicles operating outdoors often depend on GNSS, but GNSS fails in parking garages, tunnels, and other large or challenging settings. Existing feature-based localization methods often rely on vertical structures such as poles and building walls, which are sparse or absent in repetitive environments like highways and long bridges. GroundLoc addresses this by projecting LiDAR point clouds into ground-focused bird's-eye-view images, matching them against a compact 2D prior map, and correcting LiDAR odometry in 3 degrees of freedom (x, y, yaw) — using LiDAR alone, with no GNSS or IMU required.

Key Contributions

1. An open-source, 3-DOF robot localization pipeline for highly precise LiDAR-only localization using prior maps.
2. A novel 3-channel BEV image definition (intensity, slope, and z-height variance) plus demonstration that the learned keypoint network R2D2 can be successfully applied to LiDAR BEV prior-map localization.
3. A system using 3-channel raster 2D maps that consume about 4 MB/km² on average, with runtime exceeding 14 Hz on a consumer laptop using an Ouster OS2 128.
4. Validation in both single-session and multi-session settings across a variety of environments and sensors on the public SemanticKITTI and HeLiPR datasets.

Main Findings

• Outperforms state-of-the-art in multi-session localization: On the HeLiPR dataset, GroundLoc's R2D2 variant reaches an Average Trajectory Error well below 50 cm on all Ouster OS2 128 sequences while meeting online runtime requirements.
• Strong single-session accuracy on SemanticKITTI: R2D2 variant achieves an average ATE of 0.14 ± 0.03 m and average ARE of 0.21 ± 0.13°, the best ATE among compared methods, ahead of Prior Map ICP (0.18 ± 0.09 m ATE), KISS-SLAM (1.13 ± 0.88 m), Fingerprint (0.66 ± 0.79 m), and KISS-ICP (1.87 ± 1.92 m). Prior Map ICP achieves the best ARE at 0.19 ± 0.10°.
• R2D2 beats SIFT in matching, especially with sparse sensors: In single-session matching on HeLiPR KAIST04, all sensors exceeded 98% success; in multi-session matching against a map from KAIST06, success dropped for all sensors but R2D2 remained best (Aeva: 77.65% vs SIFT's 42.78%; Avia: 81.57% vs 54.15%; Ouster: 93.05% vs 85.10%).
• Large drift reduction in challenging Bridge sequences: In the Bridge scenario, GroundLoc reduced the ATE of KISS-ICP by more than a factor of 100, achieving ATEs below 1 meter, where KISS-ICP reached 137.78 m (Aeva), 169.38 m (Avia), and 13.04 m (Ouster).
• Very compact maps: GroundLoc's maps are 3-channel, 8-bit rasters at 33 cm per pixel with ZSTD compression; the densest maps (Ouster OS2 128) average 4.09 MB/km². For comparison, the Fingerprint method is reported at 33.75 MB/km², downsampled point cloud maps for ICP at 15.32 MB/km², and raw point cloud data at roughly 55.4 GB/km².
• Coarser resolution than prior work: Precise localization is achieved at 33 cm resolution, versus 15 cm or 5 cm required by other state-of-the-art methods cited.
• Cross-sensor map reuse works: Using a map built with the denser Ouster OS2 128, the Aeva Aeries II and Livox Avia localized well, with some results surpassing the same-sensor results — R2D2 Avia on Bridge improved from 0.57 m (own-sensor map, Table VI) to 0.71 m on the Ouster map in one measure, and Roundabout/Town results were competitive.
• Generalization across locations but not sensors: R2D2 trained only on KITTI-360 (Germany) transferred to a South Korean environment with minor degradation on a similar Ouster OS2 128 sensor; discrepancies increase for the Aeva Aeries II and Livox Avia, indicating difficulty transitioning between sensor models with different intensity responses and scanning patterns.
• Anomalously high errors from voxel-based ICP: Prior Map ICP failed on all Aeva sequences and the Avia Town sequence because it could not maintain map tracking, producing ATEs as high as 1195.36 m (Aeva Bridge).
• Runtime: GroundLoc processed at over 14 Hz in all experiments (the authors elsewhere state 14 Hz to 25 Hz depending on sensor) on a laptop with an Intel Core i9-13950HX and NVIDIA GeForce RTX 4090 Mobile, with KISS-ICP, GroundGrid, and BEV registration run in parallel.

Methodology in Plain English

GroundLoc builds a localization loop around three ideas:

1. Look at the ground, not the skyline. Instead of relying on vertical features like poles, the pipeline uses the GroundGrid algorithm to separate ground points from non-ground points and rasterize them into a bird's-eye-view image. Because only ground points are used, moving objects and dynamic clutter introduce far less noise.

2. Give each pixel three channels. The BEV image stores intensity (reflectance), slope (terrain gradient derived from surface normals), and z-height variance (roughness). Intensity alone was insufficient in some scenarios, so slope captures the local terrain shape and variance preserves information about static vertical structure and surface roughness. The slope channel accumulates the minimum slope observed over time rather than an average, which further suppresses dynamic-object noise.

3. Match keypoints between the live BEV and a prior map. Keypoints and 128-dimensional descriptors are extracted with either R2D2 — a small learned network with fewer than 500K parameters, trained to pick points that are repeatable and reliable — or the classical SIFT. SIFT is restricted to the intensity channel because it works on grayscale images. Descriptors are matched with an approximate nearest-neighbor KD-tree, filtered by a dynamic search radius, and the maximum consensus set is found with the Quatro estimator, which is more outlier-resistant than RANSAC. The resulting pose correction is applied only partially per frame, dampened by the number of inliers found and the current speed, to avoid oscillations from pixel quantization.

Map creation: Prior maps come from a single drive. Ground-truth poses (from an offline SLAM algorithm) are used to average BEV images at their locations, with intensity weighted by inverse variance (capped at a factor of 1000) to reduce moving-object noise. Maps are stored as compact GeoTIFF files with ZSTD compression, making them widely compatible and tiling-friendly.

Training: R2D2 was trained on BEV image pairs with synthetic distortions — random rotations in [-180°, 180°), brightness changes up to 0.5, contrast changes up to 0.3, and 192-pixel crops — plus a learning rate of 0.0001, weight decay 0.0005, and batch size 4. For SemanticKITTI it trained for 7 epochs on KITTI-360 training sequences; for HeLiPR, 15 epochs on the MulRan Sejong03 sequence followed by 3 epochs per sensor on HeLiPR's Riverside04.

Evaluation: Registration counts as successful when translation error stays under 2 m and rotation error under 5°. Localization is scored with Average Trajectory Error and Average Rotational Error, after projecting trajectories onto the xy-plane and applying Umeyama alignment (so the ARE reflects yaw only). Baselines include KISS-ICP, KISS-SLAM, a fingerprinting localization method, and Point-to-Plane ICP against voxelized (2 m voxel) point cloud maps.

Why This Matters

Impact on research: GroundLoc shows that a LiDAR-only system can compete with or beat methods that fuse GNSS and IMU, and that compact 2D raster maps can replace storage-heavy 3D point cloud maps for large-scale localization. Its 3-channel BEV definition and demonstration that a general-purpose image keypoint network (R2D2) transfers to LiDAR BEV data open a path to reusing advances in visual feature learning for LiDAR localization.

Real-world applications:

• Autonomous driving in GNSS-denied areas such as parking garages, tunnels, and long highway stretches, where the paper notes satellite localization frequently fails.
• Fleet-scale mapping where a single drive suffices to produce a prior map — helpful for covering large service areas quickly and cheaply, since maps need only a few megabytes per square kilometer.
• Multi-sensor and multi-vehicle fleets: maps built with a high-end sensor such as the Ouster OS2 128 can be consumed by cheaper or narrower-FOV platforms such as the Livox Avia, and cross-sensor localization was demonstrated.
• Long-term operation across seasons and repeated visits, where multi-session experiments on the HeLiPR dataset (recorded in August and January, introducing seasonal and long-term change) tested robustness under real change.

Industry relevance: The sub-5 MB/km² map storage, GeoTIFF compatibility, and 14+ Hz runtime on a consumer laptop fit the resource constraints of production mobile robots. The open-source release at https://github.com/dcmlr/groundloc lowers the barrier to integration and benchmarking.

Future Directions

• Sensor-to-sensor generalization: The models generalize well across locations but degrade when switching sensor models, especially the Aeva Aeries II and Livox Avia, which differ in intensity response and scanning pattern. Closing this gap could enable truly sensor-agnostic map sharing.
• Better handling of repetitive and high-speed environments: The Bridge sequence remains the hardest case, with speeds up to 90 km/h driving stronger drift. Further reducing ATE there without reliance on dense sensors is an open problem.
• Combining SIFT-style robustness with dense-sensor accuracy: SIFT matched R2D2 on the dense Ouster sensor but lost map tracking on the Bridge sequence with sparse sensors; improving the non-learning path or adapting training could make both variants reliable.
• Extending the degrees of freedom: GroundLoc only corrects x, y, and yaw, so it cannot fix pitch, roll, or z-height errors. Adding these would help on sloped terrain and remove the need for the xy-plane projection workaround applied to KISS-ICP outputs.
• Optional fusion with GNSS/IMU: The authors note the system can be combined with GNSS and IMU to increase accuracy, which remains to be explored systematically. They also attribute the Fingerprint method's higher angular error to its lack of an IMU.

Target Audience

Robotics engineers and researchers working on autonomous vehicles, mobile robot localization, and map-based navigation; teams responsible for fleet mapping and long-term map maintenance; and researchers interested in learned keypoint detection applied to LiDAR data. Readers who benefit most are those who need a practical, resource-efficient localization pipeline, or who want to understand how far LiDAR-only localization can go without GNSS, IMU, or large 3D point cloud maps. Practitioners evaluating storage, runtime, and cross-sensor map reuse trade-offs for real deployments will find the quantitative tables directly useful.

Authors’ abstract

In this letter, we introduce GroundLoc, a LiDAR-only localization pipeline designed to localize a mobile robot in large-scale outdoor environments using prior maps. GroundLoc employs a Bird's-Eye View (BEV) image projection focusing on the perceived ground area and utilizes the place recognition network R2D2, or alternatively, the non-learning approach Scale-Invariant Feature Transform (SIFT), to identify and select keypoints for BEV image map registration. Our results demonstrate that GroundLoc outperforms state-of-the-art methods on the SemanticKITTI and HeLiPR datasets across various sensors. In the multi-session localization evaluation, GroundLoc reaches an Average Trajectory Error (ATE) well below 50 cm on all Ouster OS2 128 sequences while meeting online runtime requirements. The system supports various sensor models, as evidenced by evaluations conducted with Velodyne HDL-64E, Ouster OS2 128, Aeva Aeries II, and Livox Avia sensors. The prior maps are stored as 2D raster image maps, which can be created from a single drive and require only 4 MB of storage per square kilometer. The source code is available at https://github.com/dcmlr/groundloc.