Source-linked AI summary
Range Image-based LiDAR Localization for Autonomous Vehicles
Xieyuanli Chen, Ignacio Vizzo, Thomas Läbe, Jens Behley, Cyrill Stachniss
TL;DR
Precise map-based localization is important for autonomous systems, and this paper addresses localization from 3D LiDAR scans in large outdoor maps. It compares scan-derived range images with synthetic mesh-rendered views in a Monte Carlo localization framework; experiments report reliable localization across environments and sensor types, with online operation after convergence.
Problem
The paper addresses the need for precise, reliable LiDAR-based global localization in maps, including settings where GPS cannot provide accurate localization.
Source: §I · p. 1Method
The method compares range images generated from LiDAR scans with synthetic range images rendered from a triangular-mesh map, using a new observation model in Monte Carlo localization.
Source: §I · p. 1, §III · p. 2, B. Mesh-based Map Representation · p. 3Results
The method outperforms baseline methods in location accuracy while achieving similar heading accuracy, generalizes across environments and LiDAR types, and operates online after convergence.
Source: A. Localization Performance · p. 6, B. Generalization · p. 6, C. Runtime · p. 6Takeaways & Limitations
The reported results support global localization across different environments and LiDAR sensors, with sensor-frame-rate operation after convergence.
Source: B. Generalization · p. 6, §V · p. 6
Abstract
from arXiv · showhide
Robust and accurate, map-based localization is crucial for autonomous mobile systems. In this paper, we exploit range images generated from 3D LiDAR scans to address the problem of localizing mobile robots or autonomous cars in a map of a large-scale outdoor environment represented by a triangular mesh. We use the Poisson surface reconstruction to generate the mesh-based map representation. Based on the range images generated from the current LiDAR scan and the synthetic rendered views from the mesh-based map, we propose a new observation model and integrate it into a Monte Carlo localization framework, which achieves better localization performance and generalizes well to different environments. We test the proposed localization approach on multiple datasets collected in different environments with different LiDAR scanners. The experimental results show that our method can reliably and accurately localize a mobile system in different environments and operate online at the LiDAR sensor frame rate to track the vehicle pose.
I. INTRODUCTION
The paper targets precise global localization from 3D LiDAR, where existing observation models can be limited by scanner type, handcrafted features, or environment dependence. It proposes comparing scan-derived and mesh-rendered range images in a Monte Carlo localization model.
I. INTRODUCTION
- Existing LiDAR observation models include beam-end, ray-casting, and handcrafted-feature approaches, which may be inefficient for 3D scanners or require carefully designed features.Learning-based observation models can be accurate but reliable results depend on similarity between the test environment and training environment.Source: §I · p. 1
- Range images offer a lightweight representation of rotating LiDAR scans, while compact triangular-mesh maps enable efficient rendering for large-scale global localization.The figure contrasts a mesh and its rendered range image with a LiDAR point cloud and its scan-derived range image.Source: §I · p. 1, Figure 1 · p. 1
- The proposed observation model compares range images from real LiDAR scans and synthetic mesh renderings to update particle weights for global localization.It provides (x, y, θ)-pose estimates and is evaluated for localization across environments and LiDAR scanners.Source: §I · pp. 1–2
- The paper claims reliable, accurate global localization with 3D LiDAR, generalization across environments, and compatibility with different LiDAR sensors.The authors state that experiments support these claims.Source: §I · p. 2
II. RELATED WORK
Prior work includes pose tracking that relies on GPS, particle-filter localization, and semantic or learned approaches. This paper distinguishes its LiDAR-only global-localization approach by using range information directly, without neural networks, semantics, or landmarks.
II. RELATED WORK
- Unlike pose-tracking systems that require GPS to operate on HD maps, this work addresses global localization without a pose prior from GPS or other sensors.The paper distinguishes tracking from global localization: tracking starts from a known pose, while global localization has no pose prior.Source: §II · p. 2
- Prior approaches include particle-filter methods, semantic localization using sparse maps or landmarks, and learned observation models for Monte Carlo localization.The paper notes that learning-based methods usually do not generalize well across environments or LiDAR sensors.Source: §II · p. 2
- The overview depicts scan-derived range images compared with synthetic images rendered from a mesh map, with the observation model integrated into Monte Carlo localization.The system estimates the vehicle pose posterior.Source: Figure 2 · p. 2
- The proposed method uses LiDAR range information directly, without neural networks, semantics, or landmark extraction, and does not require new training data when moving environments.The paper states that it generalizes across environments and LiDAR sensors.Source: §II · p. 2
III. OUR APPROACH
The approach projects 3D LiDAR scans into range images and compares them with synthetic views rendered from a triangular-mesh map. The resulting observation model is integrated into Monte Carlo localization.
III. OUR APPROACH
- The system compares a LiDAR-derived range image with synthetic range images rendered from the mesh map at particle locations, then uses the observation model in Monte Carlo localization.A tile map accelerates rendering and helps determine convergence; after convergence the system operates at the LiDAR sensor frame rate.Source: §III · p. 2
A. Range Image Generation
- Spherical projection maps each point to image coordinates, and the nearest 3D point populates each pixel of the vertex map.The scan points are converted through spherical coordinates to image coordinates.Source: A. Range Image Generation · p. 3
- The range image is formed from the ranges at vertex-map coordinates; the projection uses the sensor’s vertical field of view and the map’s width and height.The range of a point is r = ||p||2.Source: A. Range Image Generation · p. 3
B. Mesh-based Map Representation
The map is a triangular mesh reconstructed with Poisson surface reconstruction from oriented LiDAR point clouds. The mesh supports compact storage and fast rendering, while ground simplification further reduces its size.
B. Mesh-based Map Representation
- A triangular mesh is compact and smooth, and enables range-image rendering at the LiDAR frame rate.For each triangle, only three vertices need projection, and triangles better represent occlusion between objects than point clouds.Source: B. Mesh-based Map Representation · p. 3, Figure 3 · p. 3
- Ground-mesh simplification reduces the mesh model to about 50% of its original size.Ground vertices are contracted within voxels and averaged with adjacent vertices before invalid edges are removed.Source: B. Mesh-based Map Representation · p. 3
- Poisson surface reconstruction converts oriented LiDAR point clouds, aggregated in a global reference frame using SLAM map poses, into a globally consistent triangular mesh.The method estimates surface normals from range images and feeds the oriented point clouds to reconstruction as a whole, rather than building the map incrementally.Source: B. Mesh-based Map Representation · p. 3
C. Rendering Synthetic Range Images
The system renders synthetic range images from a triangular mesh for particle pose hypotheses, then uses Monte Carlo localization to update and resample those hypotheses. It focuses on the observation model while using a standard vehicle motion model.
C. Rendering Synthetic Range Images
- OpenGL renders a synthetic range image from the mesh at each particle pose, shading triangle surfaces with occlusion; batched rendering reduces repeated vertex reads and draw calls.Map instancing enables batches of range images to be rendered for multiple particles.Source: C. Rendering Synthetic Range Images · p. 3
D. Monte Carlo Localization
- Monte Carlo localization represents vehicle poses as particles, predicts them with a motion model, and updates their weights by comparing expected and actual observations.Particles are resampled by weight, with resampling triggered when the effective particle count falls below a threshold; the method uses a standard vehicle motion model.Source: D. Monte Carlo Localization · pp. 3–4
- The observation-model figure illustrates location likelihood at fixed heading and heading likelihood as yaw changes.Redder heatmap shades indicate higher weights.Source: Figure 4 · p. 4
E. Range Image-based Observation Model
The observation model compares the current LiDAR range image with a synthetic image rendered for each particle. It converts their image difference into a likelihood and scores the full pose state with one model.
E. Range Image-based Observation Model
- Each particle’s pose determines a rendered range image, which is compared with the current LiDAR range image to estimate its observation likelihood.The likelihood is approximated with a Gaussian distribution based on the difference or similarity between the images.Source: E. Range Image-based Observation Model · p. 4
- The chosen similarity measure is the mean absolute pixel-wise difference over valid pixels, selected as a fast-to-compute method.The passage reports that this choice is effective and generalizes across environments and LiDAR sensor types.Source: E. Range Image-based Observation Model · p. 4
- One observation model estimates likelihood over the full (x, y, θ) state space, rather than splitting location and heading likelihoods as prior work did.The paper describes the model as encoding pose hypotheses well and as elegant and fast.Source: E. Range Image-based Observation Model · p. 4
F. Tiled Map Representation
The method divides the global mesh into tiles to accelerate rendering and to identify convergence. After convergence, the tile-based approach makes runtime independent of the full environment’s size.
F. Tiled Map Representation
- The global mesh is split into tiles so rendering uses only map parts close to each particle’s position.Once all particles are localized in at most Nconv tiles, the method treats localization as converged and reduces the particle count; runtime then becomes independent of the whole environment’s size.Source: F. Tiled Map Representation · p. 4
IV. EXPERIMENTAL EVALUATION
The evaluation tests localization across datasets and LiDAR sensors, using the same MCL parameters beyond the IPB-Car tuning dataset. Baselines share the MCL framework and differ in their observation models.
IV. EXPERIMENTAL EVALUATION
- All parameters are tuned on IPB-Car and retained for experiments with other datasets and sensors.The implementation uses 100 × 100 m² tiles and reduces the particle count from 10,000 to 100 after convergence; tile size and particle reduction trade runtime against accuracy.Source: §IV · pp. 4–5
- The evaluation covers Carla, IPB-Car, MulRan, and Apollo datasets from different environments and LiDAR scanners, changing only sensor calibration parameters across datasets.Carla test sequences include objects absent from the map to simulate changing environments.Source: §IV · p. 5
- Three observation-model baselines use the same MCL framework: beam-end, histogram-based, and deep overlap-based localization.The beam-end model uses a voxelized map and can require many more particles for 3D scans; the overlap-based model estimates scan-map overlap and yaw offset with a neural network.Source: §IV · p. 5
- Fig. 5 shows the mesh map and ground-truth, overlap-based, and proposed-method trajectories for an IPB-Car run with 10,000 particles.The ground truth is black, the overlap-based result blue, and the proposed result orange.Source: Figure 5 · p. 5
A. Localization Performance
On IPB-Car, the proposed method outperforms the baseline methods in location accuracy while achieving similar heading accuracy. The section evaluates localization trajectories and success rates across particle counts.
A. Localization Performance
- The proposed method tracks the vehicle pose more accurately than baseline methods in qualitative IPB-Car trajectory comparisons.The trajectories are shown on the mesh map against ground truth in Fig. 5.Source: A. Localization Performance · p. 5
- Success rates are evaluated across five particle counts for global-localization runs on IPB-Car sequences 00 and 01.Fig. 6 shows the success rate for each observation model; a run counts as successful when location error is below 5 m at checks every 100 frames after convergence.Source: A. Localization Performance · p. 5, Figure 6 · p. 6
- The proposed method outperforms all baselines in location accuracy while achieving similar heading accuracy on IPB-Car.Tab. II reports location RMSE for successful tests with 10,000 particles and yaw-angle RMSE for converged tests at that particle count.Source: A. Localization Performance · pp. 5–6, Table II · p. 6
B. Generalization
The method localizes with five simulated LiDAR types and generalizes to MulRan and Apollo using unchanged parameters. The evaluation also reports location RMSE of 0.48 m with an 8-beam sensor.
B. Generalization
- Across five simulated LiDAR types, including an 8-beam sensor, the method achieves a location RMSE of 0.48 m with the 8-beam LiDAR.The sensors are tested for localization in the same mesh map.Source: B. Generalization · p. 6, Table III · p. 6
- The method works on MulRan and Apollo using the same parameters as on Carla and IPB-Car, across Korean and U.S. urban environments.Fig. 7 reports latitude, longitude, and heading errors relative to provided GPS locations for MulRan.Source: B. Generalization · p. 6, Figure 7 · p. 6
- The sensor-generalization results are summarized in Tab. III, which reports localization results for datasets using different sensors.Source: Table III · p. 6, B. Generalization · p. 6
C. Runtime
After convergence, the method operates online at the LiDAR sensor frame rate; experiments report reliable, accurate global localization and generalization across environments and LiDAR sensors.
C. Runtime
- After convergence, the method averages 21.8 Hz with 100 particles and 100 × 100 m2 tile maps on Carla.Before convergence, one query frame takes up to 56.7 s with 10,000 particles.Source: C. Runtime · p. 6