Research map / Architectures and efficiency / Attention and transformers
Multi-Constellation-Inspired Single-Shot Global LiDAR Localization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a single-shot global LiDAR localization method that combines scan-context coarse retrieval, a lightweight LOAM-style odometry to estimate distances to neighboring observation keyframes, and multi-sphere iterative optimization to get accurate 3-DoF (x,y,z) position.
Problem
Single-shot global LiDAR localization focuses on place retrieval and success rate, with little attention to position estimation accuracy, especially along the z-axis; registration-based two-stage methods suffer from local optima, large translations, and map-induced errors.
Approach
Inspired by multi-constellation (GNSS-style) positioning, it treats localization as triangulation: coarse retrieval yields neighboring observation points with known coordinates, LOAM-style odometry estimates distances to them, and the robot position is found by iteratively solving a multi-sphere equation system.
Key results
- Average total localization error of 0.98 m on KITTI sequences 02/05/07 (0.89 m averaged including self-collected data per abstract), versus 3.86 to 4.35 m for scan-context-based ICP/NDT/GICP/KISS-ICP baselines.
- Self-collected campus average total error 0.625 m; wilderness total error 1.13 m, about 20.89% of the second-ranked method.
- Runtime of 0.357 s per frame on KITTI and 0.214 s per frame on self-collected data; Sc-Kicp is faster but far less accurate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use multi-constellation-inspired localization instead of Sc-Icp | Global LiDAR localization · KITTI 02 | average total localization error (m): 7.95 -> 1.28 | Helped |
| Use multi-constellation-inspired localization instead of Sc-Kicp | Global LiDAR localization · KITTI 05 | average total localization error (m): 2.15 -> 0.77 | Helped |
| Use multi-constellation-inspired localization instead of Sc-Icp | Global LiDAR localization · KITTI 07 | average total localization error (m): 2.90 -> 0.90 | Helped |
| Use multi-constellation-inspired localization instead of Sc-Ndt | Global LiDAR localization · Campus 01 | average total localization error (m): 0.78 -> 0.61 | Helped |
| Use multi-constellation-inspired localization instead of Sc-Kicp | Global LiDAR localization · Campus 02 | average total localization error (m): 1.44 -> 0.64 | Helped |
| Use multi-constellation-inspired localization instead of Sc-Icp | Global LiDAR localization · Wilderness | average total localization error (m): 5.41 -> 1.13 | Helped |
Surprising findings
- The KISS-ICP-based baseline is the fastest method yet has large position errors on KITTI (around 4.3 m total).
Limitations
- Relies on coordinates of observation points from a pre-built map, GNSS, or other devices.
- Localization error remains around 1 m, and z-axis fluctuations occur under occlusion and motion distortion.
- The noise weight rho is set as a constant rather than derived from registration error variance.
Experimental setup
Datasets: KITTI sequences 02, 05, 07 self-collected campus and wilderness LiDAR sequences
Benchmarks: KITTI 02 KITTI 05 KITTI 07 self-collected campus 01 self-collected campus 02 self-collected wilderness
Metrics: average translation error standard deviation runtime per frame
Baselines: Sc-Icp Sc-Ndt Sc-Gicp Sc-Kicp
Builds on: Scan Context LOAM LIO-SAM KISS-ICP
Closest papers by meaning
- SOAR: Regression-based LiDAR Relocalization for UAVs (NeurIPS 2026 · similarity 0.84)
- DeepPointMap: Advancing LiDAR SLAM with Unified Neural Descriptors (AAAI 2024 · similarity 0.84)
- LiSA: LiDAR Localization with Semantic Awareness (CVPR 2024 · similarity 0.84)
- LightLoc: Learning Outdoor LiDAR Localization at Light Speed (CVPR 2025 · similarity 0.84)
- DiffLoc: Diffusion Model for Outdoor LiDAR Localization (CVPR 2024 · similarity 0.82)
- ForestLPR: LiDAR Place Recognition in Forests Attentioning Multiple BEV Density Images (CVPR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.