GroundLoc: LiDAR-Only Outdoor Localization

• GroundLoc is a LiDAR-only localization pipeline that converts LiDAR scans into ground-focused Bird’s-Eye View images and aligns them with compact 2D raster maps.
• It employs robust ground segmentation via GroundGrid and feature matching with R2D2 or SIFT to provide precise 3-DOF planar pose corrections in GNSS-challenged environments.
• The system achieves high localization accuracy with minimal map storage and real-time performance, supporting multi-session operations across diverse LiDAR sensors.

GroundLoc is a LiDAR-only localization pipeline for large-scale outdoor environments that localizes a mobile robot against prior maps by converting LiDAR scans into Bird’s-Eye View (BEV) images focused on the perceived ground area, then registering those BEV images to a compact raster map using R2D2, or alternatively SIFT, as the feature front end. It estimates a 3-DOF planar pose correction, is intended for settings in which GNSS may fail, visual localization may be unreliable, and dense 3D prior maps are costly to store and process, and stores prior maps as 2D raster image maps requiring only about 4 MB of storage per square kilometer on average (Steinke et al., 28 Oct 2025).

1. System definition and localization regime

GroundLoc addresses prior-map localization rather than SLAM. Its operating assumption is that a prior BEV map already exists, and that online LiDAR point clouds together with an odometric estimate can be corrected against that map. The method is explicitly LiDAR-only, and its output is a planar localization estimate in terms of $x$, $y$, and yaw rather than a full 6-DOF pose (Steinke et al., 28 Oct 2025).

The motivating failure modes are conventional in large-scale outdoor localization: GNSS can fail in tunnels, parking garages, highways, and urban canyons; visual methods often depend on stable appearance and rich vertical structure; feature-based 3D LiDAR methods often rely on poles, walls, corners, or other vertical structures that may be sparse or repetitive; dense 3D prior maps are expensive to store and register against at scale. GroundLoc responds by reformulating localization as ground-focused BEV image registration. The ground is nearly always observed by a vehicle-mounted LiDAR, using ground-only points suppresses contamination from moving vehicles and pedestrians, and roads, curbs, lane markings, slope changes, and terrain roughness provide repeatable localization cues (Steinke et al., 28 Oct 2025).

This focus on the ground area also defines the geometric scope of the method. GroundLoc is a 3-DOF system, and for sensors without full $360^\circ$ coverage it may require planar projection of odometry because pitch drift can otherwise corrupt BEV alignment. The method is therefore most naturally situated in the class of outdoor ground-vehicle localization systems that require robust planar correction over long traversals and across sessions (Steinke et al., 28 Oct 2025).

2. BEV map representation and ground-focused image formation

A central design choice in GroundLoc is the use of 3-channel, 8-bit BEV raster maps at 33 cm per pixel, stored as GeoTIFF with ZSTD compression. The representation is deliberately compact: on average, GroundLoc maps occupy about 4.09 MB/km$^2$ for Ouster OS2-128, compared with 33.75 MB/km$^2$ for the Fingerprint baseline, 15.32 MB/km$^2$ for voxelized prior point-cloud maps for ICP, and roughly 55.4 GB/km$^2$ for raw point clouds (Steinke et al., 28 Oct 2025).

GroundLoc generates these BEV images from GroundGrid, which performs ground segmentation and terrain estimation and yields a local 2D elevation representation together with a pointwise ground/non-ground partition. GroundGrid is a deterministic, non-learning LiDAR method that maintains a temporally fused local terrain map, explicitly filters points that are likely impossible observations below the ground, and was designed for robust ground estimation under slopes, curbs, non-continuous level changes, and sparse distant returns (Steinke et al., 2024).

The three BEV channels are intensity, slope, and z-height variance. They are constructed from the GroundGrid grid map $G^f$ and the sensor-centered BEV $B^f$ as follows (Steinke et al., 28 Oct 2025):

Channel	Definition	Role
Intensity	$B^f_I(u,v)=G^f_I(u,v)\cdot I_c$	Reflectivity cue
Slope	$y$0	Terrain inclination cue
Variance	$y$1	Roughness and static vertical structure cue

The single-observation slope term is defined by

$y$2

where $y$3 is the estimated surface normal at cell $y$4 and $y$5 is its z-component. GroundLoc computes this only for cells with GroundGrid ground confidence $y$6; otherwise $y$7. The use of the minimum over time for slope, rather than averaging, is intended to further reduce noise from dynamic obstacles (Steinke et al., 28 Oct 2025).

The accumulation strategy is asymmetric across channels. Intensity and variance are accumulated as a weighted average based on ground point count, whereas slope uses the minimum value. This choice reflects the claim that ground-confidence thresholding already acts as a quality filter for the slope channel, while point-count weighting is useful for stabilizing intensity and variance under LiDAR sparsity (Steinke et al., 28 Oct 2025).

3. Prior-map construction and online localization pipeline

GroundLoc prior maps can be created from a single drive. Map creation requires stored LiDAR point clouds from the mapping run and ground truth poses for each frame. For each frame, GroundLoc creates the BEV image at its true pose, projects it into the global map, and averages BEV observations at corresponding map cells. During this process, the intensity channel is weighted by the inverse of the variance to suppress noise from moving objects, with inverse variance capped at 1000 so that very small variances from low-point-count cells do not dominate the average (Steinke et al., 28 Oct 2025).

Online localization proceeds through a fixed sequence of stages: input LiDAR point cloud, GroundGrid ground segmentation and terrain estimation, 3-channel BEV image generation, keypoint and descriptor extraction, descriptor matching, dynamic positional filtering, robust registration with Quatro, and application of a damped pose correction to the odometric estimate (Steinke et al., 28 Oct 2025).

Descriptor matching is implemented with an approximate nearest neighbor KD-tree. The resulting correspondences are then filtered by positional distance using a dynamic search radius derived from previous match offsets. This exploits temporal continuity to reject implausible matches and reduce the burden on the consensus stage (Steinke et al., 28 Oct 2025).

For robust BEV-to-map registration, GroundLoc uses Quatro with parameters $y$8 and $y$9. Quatro was selected because it finds a maximum consensus set heuristically, is more resistant to outliers than RANSAC, and remains fast in typical operating conditions. When the number of matching keypoints becomes very large, GroundLoc applies a simple hysteresis mechanism that dynamically scales the maximum allowed feature distance and maximum keypoint number, because Quatro can degrade when there are too many matching keypoints (Steinke et al., 28 Oct 2025).

The pose correction is not applied in full. To prevent sudden jumps due to occasional false matches, GroundLoc uses damped and capped updates. For the $360^\circ$0-component,

$360^\circ$1

where $360^\circ$2 is the estimated offset, $360^\circ$3 is the dampening factor, and $360^\circ$4 is the correction cap. The cap is defined by

$360^\circ$5

with $360^\circ$6 the number of Quatro inliers, $360^\circ$7 the vehicle speed in m/s, and $360^\circ$8 a correction factor. The same formulation is used analogously for $360^\circ$9 and yaw. The paper uses $^2$0, $^2$1 for R2D2, and $^2$2 for SIFT (Steinke et al., 28 Oct 2025).

4. Keypoint front ends: R2D2, SIFT, and sensor adaptation

GroundLoc supports two interchangeable front ends. The preferred learned option is R2D2, a compact CNN with fewer than 500K parameters that performs both keypoint detection and descriptor extraction. R2D2 identifies keypoints from the product of repeatability and reliability. In GroundLoc’s BEV images, repeatability tends to highlight road structure, curbs, and lane markings, while reliability tends to suppress occluded areas, noisy vegetation, and dark or low-texture regions. The alternative, SIFT, is a training-free option that is restricted to the intensity channel because it operates on grayscale images only (Steinke et al., 28 Oct 2025).

R2D2 is explicitly trained for LiDAR-BEV localization. Training pairs are generated from BEV images and corresponding map crops using ground-truth poses, with image pairs separated by at least 1 meter to avoid overrepresenting stationary frames. Augmentations include random rotation in $^2$3, noise, intensity variations, brightness changes, contrast changes, and random cropping with crop-size parameter 192 px. For SemanticKITTI, the network is trained for 7 epochs on KITTI-360 training sequences. For HeLiPR, it is trained for 15 epochs on MulRan Sejong03, then fine-tuned for 3 epochs on HeLiPR Riverside04 for each sensor, with learning rate $^2$4, weight decay $^2$5, and batch size 4 (Steinke et al., 28 Oct 2025).

The paper reports that GroundLoc accommodates different LiDAR sensors through a combination of sensor-specific GroundGrid parameters and sensor-specific BEV normalization factors. The normalization factors $^2$6 are given as follows: HDL-64E $^2$7, OS1-64 $^2$8, OS2-128 $^2$9, Aeva Aeries II $^2$0, and Livox Avia $^2$1 (Steinke et al., 28 Oct 2025).

For sensors without $^2$2 coverage, GroundLoc projects odometry into the plane to suppress pitch-drift artifacts. The projected update is

$^2$3

where $^2$4 is the projected 2D position at frame $^2$5, $^2$6 is the previous 2D position, $^2$7 is the 3D odometry increment, and $^2$8 is its planar component (Steinke et al., 28 Oct 2025).

5. Empirical evaluation, datasets, and reported performance

GroundLoc is evaluated on SemanticKITTI for single-session localization and HeLiPR for multi-session localization across different LiDAR sensors. The reported metrics are Average Trajectory Error (ATE) and Absolute Rotational Error (ARE). Matching quality is also evaluated, with a registration counted as successful if translation error < 2 m and rotation error < $^2$9 (Steinke et al., 28 Oct 2025).

On single-session matching using HeLiPR KAIST04, both front ends perform strongly, but R2D2 is usually slightly stronger. For Aeva, SIFT reaches 0.69 m / $^2$0 / 98.46%, whereas R2D2 reaches 0.60 m / $^2$1 / 99.33%. For Avia, SIFT gives 0.25 m / $^2$2 / 99.70% and R2D2 gives 0.30 m / $^2$3 / 99.80%. For Ouster, SIFT gives 0.51 m / $^2$4 / 99.66% and R2D2 gives 0.43 m / $^2$5 / 99.86% (Steinke et al., 28 Oct 2025).

On multi-session matching, the difference becomes much larger. Using a map from KAIST06 and queries from KAIST04 under August-to-January change, Aeva SIFT degrades to 11.25 m / $^2$6 / 42.78%, while Aeva R2D2 reaches 4.41 m / $^2$7 / 77.65%. For Avia, SIFT gives 8.05 m / $^2$8 / 54.15%, while R2D2 gives 3.47 m / $^2$9 / 81.57%. For Ouster, SIFT gives 2.97 m / $^2$0 / 85.10%, while R2D2 gives 1.87 m / $^2$1 / 93.05%. This suggests that R2D2 is the more robust front end in multi-session and sparse-sensor regimes (Steinke et al., 28 Oct 2025).

On SemanticKITTI single-session localization, GroundLoc attains the best average ATE among the compared methods. GroundLoc R2D2 reports $^2$2 m ATE and $^2$3 ARE; GroundLoc SIFT reports $^2$4 m and $^2$5. The corresponding baselines are KISS-ICP with $^2$6 m, KISS-SLAM with $^2$7 m, Prior Map ICP with $^2$8 m, and Fingerprint with $^2$9 m (Steinke et al., 28 Oct 2025).

On HeLiPR multi-session localization, the strongest headline result is that GroundLoc achieves ATE well below 50 cm on all Ouster OS2-128 sequences. With R2D2, the reported results are 0.42 / 0.48 on Roundabout, 0.40 / 0.46 on Town, and 0.39 / 0.41 on Bridge, in ATE / ARE form. For Aeva Aeries II, GroundLoc R2D2 gives 1.03 / 0.61, 1.17 / 0.81, and 0.80 / 0.87 across the same three sequences. For Livox Avia, GroundLoc R2D2 gives 0.57 / 0.66, 0.54 / 0.68, and 0.57 / 1.11; in contrast, SIFT loses tracking on Avia Bridge with 60.00 / 2.22 (Steinke et al., 28 Oct 2025).

The runtime is reported as more than 14 Hz in all experiments, with an overall 14–25 Hz frame-rate range depending on the sensor, on a laptop with Intel Core i9-13950HX and NVIDIA GeForce RTX 4090 Mobile. Efficiency is attributed to the 2D raster map representation, approximate nearest-neighbor matching, dynamic match filtering, Quatro registration, and the parallel execution of KISS-ICP, GroundGrid, and BEV registration (Steinke et al., 28 Oct 2025).

6. Position within ground-centric localization research

GroundLoc belongs to a broader class of ground-centric localization methods, but its modality and map representation are specific. A conceptually distinct predecessor is Micro-GPS, which performs global localization from a downward-facing camera by indexing ground texture and using SIFT, PCA-compressed descriptors, precise Hough-style voting, and RANSAC on a pre-registered texture map (Zhang et al., 2017). GroundLoc, by contrast, is LiDAR-only, operates on 3-channel BEV raster maps, and estimates planar pose by feature-based BEV image registration rather than close-range image matching of texture patches (Steinke et al., 28 Oct 2025).

The system’s ground-centric emphasis is also closely tied to GroundGrid, whose terrain estimation and ground segmentation make it possible to construct stable BEV channels from intensity, slope, and variance. This suggests that GroundLoc is not merely a feature-matching front end but a compound design in which ground modeling and localization are tightly coupled through the BEV representation (Steinke et al., 2024).

Its limitations are explicit. GroundLoc is 3-DOF only; it does not estimate pitch, roll, or height. It requires a prior map, and map creation requires ground truth poses from a high-quality offline process. The paper reports good location generalization but weaker sensor-model generalization, attributing this to differences in intensity response and scanning patterns. SIFT is notably fragile in difficult multi-session sparse-sensor settings. The authors also note that for some dense Ouster conditions, Prior Map ICP can still obtain the best raw accuracy, and that the multi-session ground truth may itself limit measurable accuracy on some Ouster sequences (Steinke et al., 28 Oct 2025).

These constraints place GroundLoc in a specific niche: a LiDAR-only, map-based, online, ground-focused outdoor localization pipeline that trades full 6-DOF estimation and dense 3D map registration for compact raster maps, strong multi-session robustness, and broad sensor support. In that niche, it demonstrates that highly accurate large-scale outdoor localization does not require dense 3D prior maps or a reliance on vertical landmark structure, provided that the ground area is represented in a sufficiently informative BEV form (Steinke et al., 28 Oct 2025).