PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Global LiDAR Localization: Challenges, Advances and Open Problems

Huan Yin, Xuecheng Xu, Sha Lu, Xieyuanli Chen, Rong Xiong, Shaojie Shen, Cyrill Stachniss, Yue Wang

arXiv:2302.07433v5cs.RO

TL;DR

Global LiDAR localization requires estimating robot pose on a map across challenging, changing environments, while existing survey coverage had not explicitly integrated LiDAR place recognition and pose estimation. This survey synthesizes probabilistic formulations, map and localization methods, and cross-robot settings across three themes, reporting strong place-recognition performance but continuing limitations in generalization, efficiency, and long-term evaluation.

  • Problem

    Global localization is needed for mobile robots, but LiDAR localization spans diverse maps, measurements, pose-estimation strategies, and cross-robot settings that require comprehensive synthesis.

  • Method

    The survey organizes global LiDAR localization around probabilistic formulation, map representations, place retrieval with pose estimation, sequential measurements, cross-robot localization, and open problems.

  • Results

    More than 95% Recall@1 has been reported for several LiDAR place-recognition approaches, while coupled methods can improve pose-estimation accuracy and efficiency through better initial guesses.

  • Takeaways & Limitations

    Global LiDAR localization methods must be selected and customized according to environments, maps, required pose accuracy, and application scenarios.

  • Takeaways & Limitations

    Long-term evaluation remains insufficient because conventional metrics mainly emphasize short-term place-retrieval performance.

Abstract

from arXiv · show

Knowledge about the own pose is key for all mobile robot applications. Thus pose estimation is part of the core functionalities of mobile robots. Over the last two decades, LiDAR scanners have become the standard sensor for robot localization and mapping. This article aims to provide an overview of recent progress and advancements in LiDAR-based global localization. We begin by formulating the problem and exploring the application scope. We then present a review of the methodology, including recent advancements in several topics, such as maps, descriptor extraction, and cross-robot localization. The contents of the article are organized under three themes. The first theme concerns the combination of global place retrieval and local pose estimation. The second theme is upgrading single-shot measurements to sequential ones for sequential global localization. Finally, the third theme focuses on extending single-robot global localization to cross-robot localization in multi-robot systems. We conclude the survey with a discussion of open challenges and promising directions in global LiDAR localization. To our best knowledge, this is the first comprehensive survey on global LiDAR localization for mobile robots.

1 Introduction

Global LiDAR localization estimates a robot’s pose on a given map without an initial pose, a larger search problem than local tracking. The survey formulates this probabilistically and organizes methods by measurement sequence, coupling of retrieval and pose estimation, and robot configuration.

  • Problem formulation: Global localization estimates a robot pose from scratch on a given map, unlike pose tracking, so its solution space is generally larger.The survey frames the problem using a map, measurements, poses, likelihood, and prior information.
  • Motivation: GNSS, emitters, and external markers provide global positioning, but their usefulness can be limited by signal conditions or infrastructure requirements.These constraints motivate onboard-sensor approaches that avoid modifying the environment.
  • Motivation: LiDAR’s direct range measurements are more robust to illumination and appearance changes than camera images, supporting global localization in changing, large-scale environments.The survey notes LiDAR’s progression from 2D sparse scans to 3D point clouds and its increasing adoption in robotics.
  • Survey scope: Single-shot localization estimates one global pose from one LiDAR point cloud using maximum likelihood, with place recognition and pose estimation as its principal method categories.Place recognition retrieves candidate places, while pose estimation supplies a fine-grained metric pose; their coupling varies across surveyed approaches.
  • Survey scope: Sequential localization replaces one measurement with a scan or submap stream, using temporal connections and optionally odometry to estimate a trajectory or final pose.The survey covers both batch fusion of single-shot results and recursive filtering, while practical systems may track multiple hypotheses.
  • Survey organization: The survey reviews maps, mainstream single-shot and sequential methods, cross-robot localization, and open problems, extending coverage to topics not explicitly reviewed in prior surveys.Its organization follows three themes: retrieval with local estimation, sequential measurements, and cross-robot localization.

2 Maps for Global Localization

The survey presents keyframe-based submaps, global feature maps, and global metric maps as foundational representations for global LiDAR localization. These choices trade retrieval and maintenance convenience, feature stability, geometric precision, and resource demands.

  • Map taxonomy: The survey classifies general-use localization maps into keyframe-based submaps, global feature maps, and global metric maps.It introduces these three map structures before reviewing concrete localization methods.
  • Keyframe-based submaps: Keyframe-based submaps pair robot poses with aligned local submaps and connections, discretizing pose space for place retrieval.Each keyframe can provide a global descriptor and may include metric points or grids for geometric registration.
  • Keyframe-based submaps: Keyframe spacing controls map size and localization risk: fewer keyframes reduce lightweight-navigation requirements but increase the chance of localization failure.The survey identifies keyframe resolution and pose distance as important practical design factors.
  • Global feature maps: Global feature maps represent environments with sparse local feature points, but stable feature generation and maintenance are challenging and may require costly HD-map infrastructure or strong feature extraction.Feature correspondence can support relative transformation estimation when feature points contain high-dimensional information.
  • Global metric maps: Global metric maps provide dense geometric representations such as points, grids, voxels, or meshes, but their size can burden resource-constrained robots.Downsampling or compression may reduce localization performance, motivating sparse features or submap partitioning.
  • Implicit representations: Implicit representations, including probability-density and learning-based forms, are increasingly used because they can require fewer parameters than explicit representations.The survey describes implicit maps as an emerging alternative alongside the three principal map categories.

3 Single-shot Global Localization: Place Recognition and Pose Estimation

Single-shot global localization estimates a robot pose from one LiDAR point cloud, using place recognition and pose estimation with increasing degrees of coupling. The survey organizes methods into four categories spanning retrieval-only systems, coarse-to-fine pipelines, coupled approaches, and direct global pose estimation.

  • Overview: Single-shot global localization estimates pose from one LiDAR point cloud, typically retrieving the most probable place from keyframe-based submaps.Global descriptors should distinguish different places while remaining similar for nearby places.
  • Place Recognition Only: Place recognition alone returns a retrieved keyframe as the estimated pose, without a separate pose-estimation stage.
  • Place Recognition Followed by Local Pose Estimation: Place recognition followed by local pose estimation first retrieves a place and then estimates the robot pose with a customized estimator.
  • Coupled and One-stage Methods: Coupled methods share representations between place recognition and pose estimation, while one-stage methods estimate the global pose directly on a global map.
  • Map Organization: Methods in Sections 3.1–3.3 generally use keyframe-based submaps, whereas Section 3.4 generally uses a global feature map or metric map.

3.1 Place Recognition Only

Place recognition-only methods retrieve locations from pre-built keyframe-based maps using global descriptors. Neural descriptors have achieved strong retrieval performance, but retrieved places provide limited metric pose information and compressed representations lose geometric detail.

  • Method: Place recognition-only approaches retrieve the most similar place from a pre-built keyframe-based map using global descriptors.Descriptor extraction is challenging because LiDAR point clouds are textureless, irregular, and sometimes unevenly dense.
  • Descriptor Families: Descriptor pipelines include dense points or voxels, sparse segments, and projections of 3D point clouds into lower-dimensional representations.
  • Evaluation: Place recognition-only systems output retrieved places, so evaluation emphasizes retrieval metrics such as precision-recall curves and F1 score.Translation precision depends on keyframe resolution, while rotation estimation is not considered in this setting.
  • Limitation: Global descriptors compress raw LiDAR data and may lose information, making them suitable for nearest-neighbor retrieval but unsuitable for geometric pose estimation.

3.2 Place Recognition Followed by Local Pose Estimation

This category combines global place retrieval with independent local registration in a coarse-to-fine pipeline. Registration methods estimate precise transformations but face challenges from unknown correspondences, local minima, outliers, low overlap, and initialization sensitivity.

  • Pipeline: The pipeline first retrieves a place from keyframe-based submaps, then matches the input LiDAR cloud to map data attached to that keyframe.
  • Registration Formulation: Point cloud registration estimates a transformation by minimizing an error function between source measurements and target map points.Correspondence-based and correspondence-free methods differ in whether they use explicit spatial data associations.
  • Correspondence-based Methods: ICP alternates correspondence search and optimization, but its variants can fall into local minima and become unsuitable for global registration.Go-ICP provides a global ICP solution through branch-and-bound, though it can be time-consuming when the pose space is large.
  • Robust Estimation: Correspondence-based registration is constrained by outliers and low overlap, motivating robust estimators and application-specific assumptions such as urban-scene structure.
  • Correspondence-free Methods: Correspondence-free methods align point clouds through feature similarity, but iterative locally convergent solvers remain sensitive to initialization.
  • Summary: Combining place recognition with local pose estimation yields a complete coarse-to-fine global localization framework, while global registration can provide more reliable local poses between measurements and retrieved places.

3.3 Pose Estimation-coupled Place Recognition

Pose estimation-coupled place recognition shares representations to retrieve places and estimate metric pose more tightly than a two-step pipeline. These methods can provide 3-DoF, yaw-only, or 6-DoF outputs, but accurate poses may still require local registration refinement.

  • Coupled Design: Coupled methods share feature representations or processing between place recognition and pose estimation, reducing separation between the two stages.
  • 3-DoF Methods: Scan-context methods represent 3D clouds with azimuthal and radial bins, while later variants improve discriminability, invariance, and similarity computation.
  • Invariance: Standard scan-context descriptors are rotation-invariant for retrieval but cannot provide translation invariance because of egocentric modeling.Later descriptors introduce translation estimation or joint rotational and translational invariance.
  • 6-DoF Methods: Projection-based and visual-inspired approaches use BEV, spherical, or other 2D representations, while 3D feature and graph methods support place retrieval and 6-DoF pose estimation.
  • Outputs and Evaluation: Coupled methods can produce retrieved places together with 3-DoF, yaw-only, or full 6-DoF poses, evaluated using retrieval metrics and pose errors.
  • Potential Advantages: Shared representations may reduce map size and enable geometric verification of retrieved places through pose-estimation results.
  • Pose Refinement: When accurate pose is required, local registration remains necessary; coupled methods can provide its initial guess and reduce runtime and metric errors in reported LCDNet results.

3.4 One-stage Global Pose Estimation

One-stage global pose estimation matches LiDAR data directly against global representations, avoiding explicit place recognition in some approaches. Feature-based methods exploit segments or semantic structures, while deep regression methods predict poses end to end but remain limited in large-scale relocalization and metric accuracy.

  • 3.4 One-stage Global Pose Estimation: One-stage methods estimate global poses without separating place recognition from pose estimation.They directly match against global maps or regress the pose from LiDAR data.
  • 3.4.1 Feature-based Matching: SegMatch segments dense LiDAR maps, extracts segment descriptors, classifies matches, and applies RANSAC for 6-DoF pose estimation.SegMap extends the handcrafted descriptors with deep neural networks, while SemSegMap integrates visual information.
  • 3.4.1 Feature-based Matching: Segment-based methods align point-cloud segments using low-dimensional descriptors and can support loop closing and online pose tracking.Their segment-level operation differs from conventional keypoint-based localization.
  • 3.4.1 Feature-based Matching: Segment matching is less efficient for fast relocalization with sparse scans and can fail in featureless or highly repetitive environments.These limitations follow from the need to accumulate dense clouds and rely on segmentation results.
  • 3.4.1 Feature-based Matching: Feature-based one-stage approaches suit small-area loop closure but make from-scratch relocalization difficult when only partial local features are available in large maps.Their effective solution space is smaller for local loop-closure applications.
  • 3.4.2 Deep Regression: One-stage deep regression is fully data-driven, but its metric estimation is not competitive with conventional two-stage methods and still raises interpretability and generalization issues.PointLoc uses an attention-aided PointNet-style architecture for 6-DoF pose regression.

4 Sequential Global Localization

Sequential global localization extends single-shot LiDAR localization by using multiple measurements and sequence information. The survey distinguishes place matching on keyframe submaps from metric localization on geometric maps, with batch and recursive estimation frameworks spanning several back ends.

  • 4 Sequential Global Localization: Sequential global localization uses multiple LiDAR measurements because single-shot methods face large map-to-scan scale differences and challenging scenes.Sequential information can be handled through retrieval, optimization, or filtering.
  • 4 Sequential Global Localization: Sequential place matching fuses single-shot global localization results on keyframe-based submaps, whereas sequential metric localization fuses non-global measurements on metric maps.Metric localization generally requires a back-end state estimator for multi-hypothesis tracking.
  • 4 Sequential Global Localization: Batch methods process sequential information jointly to estimate a trajectory, while recursive filtering estimates pose incrementally through Bayesian filtering or similar techniques.The distinction concerns how sequential information is incorporated into global pose estimation.
  • 4.1 Sequential Place Matching: Sequential place-matching methods improve retrieval by combining sequence information with LiDAR descriptors and, in some designs, spatial-temporal transformer fusion.Reported strategies include coarse-to-fine matching and particle-aided matching for large-scale environments.
  • 4.2 Sequential Metric Localization: Metric sequential localization uses geometric maps and back ends such as MCL, multiple-hypothesis tracking, and factor graphs to estimate global poses.MCL represents multimodal robot states with particles, while factor-graph methods optimize batches of odometry and landmark information.
  • 4.2 Sequential Metric Localization: LiDAR-based differentiable particle filtering remains an open direction because particle filtering is mature for LiDAR localization but no LiDAR-based differentiable state estimator is currently reported.The survey identifies differentiable state estimation as promising in the era of big data.

5 LiDAR-aided Cross-robot Localization

Cross-robot localization extends global LiDAR localization across robots that map and localize on one another’s maps. The survey highlights recognition, graph optimization, and robustness to false loop closures as central components and challenges.

  • 5 LiDAR-aided Cross-robot Localization: Cross-robot localization lets one robot estimate its pose on a map built by another robot, with either robot potentially performing either role.The topic extends the survey’s single-robot focus to multi-robot systems.
  • 5.1 LiDAR-aided Multi-robot System: Multi-robot systems may use distributed architectures or centralized servers, while this section emphasizes incremental keyframe-based cross-robot localization.Communication bandwidth and computation efficiency are identified as broader system concerns outside the paper’s main focus.
  • 5.1 LiDAR-aided Multi-robot System: Cross-robot place recognition is needed when robots begin tasks in different locations, including large-scale search-and-rescue settings.Approaches use descriptors such as NetVLAD-based thumbnails or LiDAR scan context for place recognition.
  • 5.1 LiDAR-aided Multi-robot System: Inter-robot loop-closure detection enables cross-robot localization in large-scale environments, but false positives can create inconsistent graph links and destabilize optimization.The survey notes that no LiDAR loop-closure detector provides perfect loop closures without false positives.
  • 5.2 Cross-robot Back-end: Robust back ends address false loop closures either by rejecting outliers before optimization or by reducing their influence during graph optimization.Examples include RANSAC, SCGP, maximum-set estimation, switchable constraints, and robust iterative methods.
  • 5 LiDAR-aided Cross-robot Localization: Cross-robot localization remains a promising direction involving odometry, loop-closure detection, and robust back-end estimation.The survey relates it to crowd-sourced mapping while noting that additional topics lie beyond its scope.

6 Open Problems

Open problems span method selection, evaluation, long-term robustness, public datasets, multimodal localization, and system integration. The survey emphasizes that no single method suits all environments, maps, and pose-accuracy requirements.

  • Method Selection: No single global LiDAR localization method handles all environments, maps, and required pose accuracies.The survey recommends tailoring systems to application requirements.
  • Evaluation Difference: Place-retrieval metrics include Recall@1%, precision-recall curves, localization probability, and F1, while pose estimation uses translation and rotation errors.The survey notes that metric choice depends on the practical task, such as loop closure or re-localization.
  • Evaluation Difference: Pose estimation metrics may be more meaningful than place retrieval because downstream planning and control require accurate pose states.The survey identifies evaluation using a common 6-DoF pose metric as a future direction for LiDAR localization.
  • Long-term Evaluation: Long-term evaluation remains open because existing Forgetting scores derive from place-recognition precision-recall results and may not suit metric localization.The survey calls for novel metrics for long-term global localization.
  • Public Datasets: Public datasets differ in scenarios, challenges, and viewpoint diversity, requiring dataset selection that matches the intended evaluation conditions.Boreas is described as a multi-session urban dataset containing LiDAR and radar range sensing.
  • Multiple Modalities: Multimodal fusion can improve place recognition, but it requires additional learning techniques or training data.Examples combine visual, LiDAR, radar, or satellite modalities through shared representations or fusion modules.

7 Conclusion

This survey aggregates advanced knowledge on global LiDAR localization and organizes it around place retrieval and pose estimation, sequential measurements, and cross-robot localization. It also highlights continuing research directions and the need to tailor localization systems to application requirements.

  • Survey Scope: The survey reviews global LiDAR localization through probabilistic formulation, application scopes, and three themes covering retrieval, sequential localization, and cross-robot systems.The themes connect global place retrieval with local pose estimation, extend single-shot measurements to sequences, and broaden localization across robots.
  • Open Directions: Global LiDAR localization remains an active problem with promising research directions and valuable questions about integration into navigation systems.The survey points to system architecture and operating environments as additional research considerations.
  • Open Directions: Users should tailor global localization systems to their specific requirements because robotics applications are often case-specific.
Loading 2302.07433v5…