PaperlayerView exact revision on arXiv

Source-linked AI summary

Chain-SLAM: Globally Consistent Backend for Multi-Session LiDAR SLAM via Chained Loop Closure

Zhiheng Li, Xinhao Liu, Juexiao Zhang, Yongqing Liang, Chen Feng

arXiv:2609.12221v1cs.RO

TL;DR

Large-scale LiDAR SLAM still faces challenges maintaining consistency when maps are collected across multiple sessions and over long horizons. Chain-SLAM aligns and reuses maps online, propagating validated loop-closure constraints through an adjacency graph; experiments report improved accuracy and consistency across large-scale datasets.

  • Problem

    Large-scale LiDAR SLAM has difficulty maintaining consistency over long spatial and temporal horizons when integrating maps collected across multiple sessions.

    Source: §I · p. 1

  • Method

    Chain-SLAM performs online multi-session map alignment and reuse, propagating inter-session geometric constraints through an adjacency graph of connected keyframes.

    Source: §I · p. 1, A. Multi-Session SLAM and Map Merging · p. 2

  • Results

    Experiments report improved trajectory accuracy, geometric consistency, and viewpoint alignment over existing methods, with minimal cross-platform tuning and real-time performance.

    Source: §I · p. 2, §V · p. 8

  • Takeaways & Limitations

    The results support Chain-SLAM’s use for robust long-term mapping and localization in large-scale multi-session environments.

    Source: §V · p. 8

    Map merging relies on GNSS to propose cross-session correspondences, limiting direct deployment in GNSS-denied environments.

    Source: §V · p. 8

Abstract

from arXiv · show

Maintaining consistency over long spatial and temporal horizons remains a fundamental challenge in large-scale LiDAR SLAM, particularly when integrating maps collected across multiple sessions. We present Chain-SLAM, a LiDAR SLAM backend enabling online multi-session map alignment and reuse with global consistency at large scale. We implement a chained loop closure mechanism that efficiently propagates geometric constraints across inter-session keyframes through an adjacency graph, enabling robust long-horizon consistency triggered by reliable short-horizon loop closures. The system initializes inter-session alignment with GNSS-proximity place recognition, then performs on-the-fly loop closure detections and joint optimization of loaded maps and newly acquired trajectories within a unified factor graph, maintaining both inter- and intra-session geometric consistency without dynamic object removal, and cross-platform robustness with minimal hyperparameter tuning. Experimental results show improved trajectory accuracy and robust multi-session integration on large-scale datasets. We release our source code to support reproducible research in large-scale multi-session LiDAR SLAM. Project site: https://ai4ce.github.io/Chain-SLAM/

I. INTRODUCTION

Large-scale, multi-session LiDAR SLAM lacks real-time map reuse and incremental global optimization, despite prior maps offering useful priors across changing environments. Chain-SLAM combines online map integration with chained loop closure, and experiments report improved accuracy and consistency across large-scale datasets.

I. INTRODUCTION

  • Long spatial and temporal horizons bring environmental changes during revisits, while earlier-session maps can provide valuable priors for new sessions.The paper identifies a lack of open-source systems supporting real-time multi-session mapping or on-the-fly map reuse.Source: §I · p. 1
  • Existing SLAM lacks multi-session deployment or map reuse, while post-processing map merging cannot use prior-session information during operation and struggles to scale across many revisits.Large-scale operation also faces platform-dependent performance and reliance on high-fidelity sensors, high-rate measurements, or extensive tuning.Source: §I · p. 1
  • Chain-SLAM jointly optimizes trajectories across long horizons and supports loading maps, GNSS-based online alignment, incremental map reuse, and chained loop closure through a connectivity graph.The system uses scan-to-submap registration for loop closing and submap-to-submap registration for map merging, with minimal hyperparameter tuning across platforms.Source: §I · p. 1
  • Experiments on MARS and NCLT cover over 20 hours and report improved trajectory accuracy, geometric consistency, and viewpoint alignment, with minimal cross-platform tuning and real-time performance.The evaluations span diverse sensing and motion conditions.Source: §I · pp. 1–2

II. RELATED WORKS

Prior multi-session SLAM often merges completed maps after operation, while long temporal gaps challenge cross-session recognition and registration. Chain-SLAM integrates maps online and propagates reliable correspondences through connected keyframes to add constraints without exhaustive search.

A. Multi-Session SLAM and Map Merging

  • LT-Mapper and LAMM merge independently built maps primarily as post-processing, limiting immediate reuse and complicating large-scale optimization as inter-session constraints grow.Multi-agent systems, also align maps jointly but typically assume short temporal gaps and minor environmental changes.Source: A. Multi-Session SLAM and Map Merging · p. 2
  • Visual-inertial systems ORB-ATLAS and ORB-SLAM3 support online submap merging and joint optimization after cross-map correspondence or relocalization.These systems provide an online map-reuse precedent in a different sensor domain.Source: A. Multi-Session SLAM and Map Merging · p. 2
  • Chain-SLAM treats a loaded map as a base map and the incoming session as a new submap, then relocalizes, merges, and jointly optimizes them in a unified pose graph.Chained loop closure adds inter-session constraints through connected keyframes, reducing correspondence-search cost and avoiding place-recognition false negatives without explicit dynamic-object removal.Source: A. Multi-Session SLAM and Map Merging · p. 2

B. Sequential and Multi-Constraint Data Association

  • Sequential VPR methods aggregate evidence over time or model spatio-temporal persistence [13],, while panoramic approaches exploit broader context and self-supervision.These approaches motivate robustness to appearance and viewpoint variation.Source: B. Sequential and Multi-Constraint… · p. 2
  • Consistency-maximization methods, prune inconsistent associations, whereas Chain-SLAM expands geometric validation from reliable seed matches by retrieving connected keyframes.The resulting constraints crossvalidate sessions: existing structures stabilize the new trajectory, while new observations correct the loaded map and reduce drift, with minimal overhead.Source: B. Sequential and Multi-Constraint… · p. 2

III. METHOD

Chain-SLAM combines a tightly coupled LiDAR–inertial frontend with a factor-graph backend that jointly optimizes loaded maps and new trajectories. Incremental optimization updates the graph as new factors arrive to maintain global alignment.

A. System Overview

  • The system uses a tightly coupled LiDAR–inertial odometry frontend and a backend extending Fast LIO SAM with chained loop closure and map merging in a unified factor graph.Previously constructed maps can be loaded and jointly optimized with all sessions.Source: A. System Overview · pp. 2–3
  • The factor graph represents keyframe poses with unary priors anchoring trajectory starts and binary between-factors encoding relative pose constraints.Between-factors include odometry and loop closures, with geometric registration connecting non-consecutive keyframes.Source: A. System Overview · p. 3
  • The frontend connects consecutive keyframes with odometry factors, while geometric-registration loop closures connect non-consecutive keyframes.The optimal trajectory is obtained by minimizing the graph objective.Source: A. System Overview · p. 3
  • iSAM2 incrementally updates only affected graph portions when new factors are added, propagating pose corrections throughout the graph.The stated purpose is to maintain global geometric consistency.Source: A. System Overview · p. 3

B. Loop Closure Detection and Registration

Loop closures are proposed by spatial and temporal filtering and accepted through scan-to-submap ICP checks. Chain-SLAM then uses an adjacency graph to retrieve connected keyframes and add validated inter-session constraints beyond each direct match.

B. Loop Closure Detection and Registration

  • Candidates come from keyframes within a spatial radius and beyond a minimum temporal interval; ICP registers the current scan against the matched past keyframe’s submap.ICP estimates the rigid transformation by minimizing mean squared alignment error.Source: B. Loop Closure Detection and… · p. 3

T MSE(T),

  • Accepted loop closures require ICP convergence and error eij≤τ, and their factor covariance scales with that error to assign higher confidence to consistent alignments.These factors reduce drift and enforce geometric consistency across revisited regions.Source: T MSE(T) · p. 3

C. Chained Loop Closure

  • Direct closures correct drift between matched keyframes, but residual inconsistencies can remain across large multi-session maps, motivating chained loop closure.The method propagates constraints beyond a single direct correspondence while seeking to preserve local accuracy.Source: C. Chained Loop Closure · p. 3
  • Accepted intra- and inter-session loop closures form an adjacency graph whose connected components expand as loaded sessions add inter-session edges.Traversing components retrieves cross-session correspondences without repeated global place recognition or spatial search.Source: C. Chained Loop Closure · pp. 3–4
  • After a direct match, breadth-first search identifies connected loaded keyframes; ICP validates each candidate and adds a chained factor when it converges within the error threshold.Chained closures apply only between current-session and loaded-map keyframes, avoiding redundant constraints on locally consistent intra-session odometry.Source: C. Chained Loop Closure · p. 4

D. Map Representation and Multi-Session Integration

Maps are serialized as indexed keyframes, scans, poses, loop closures, and scene ranges, then restored as connected factor-graph components. GNSS-proximity matching initializes alignment, after which loop closures connect loaded and incoming sessions for joint optimization.

D. Map Representation and Multi-Session Integration

  • Saved maps preserve indexed keyframes, poses, LiDAR scans, loop-closure constraints with covariances, and scene ranges for restoration into the global factor graph.Restored maps form connected graph components with their original covariances and session membership.Source: D. Map Representation and… · p. 4
  • Chained loop closure searches reachable loaded keyframes from a matched keyframe, validates candidates with ICP, and adds accepted loop-closure factors.The procedure avoids self-matching the matched frame and derives factor covariance from ICP fitness.Source: D. Map Representation and… · p. 4
  • Incoming keyframes receive new indices and odometry factors, while inter-session alignment and loop closures connect their component to loaded maps for unified optimization.The resulting graph supports consistent multi-session integration.Source: D. Map Representation and… · p. 4
  • GNSS proximity search retrieves candidate keyframes, and FPFH with RANSAC estimates an initial inter-session transformation.The RANSAC fitness is the inlier-correspondence count divided by target-submap points, and the transformation is accepted only above a predefined threshold.Source: D. Map Representation and… · p. 4
  • The first accepted transformation aligns the current trajectory with the loaded map, then the factor graph is rebuilt with transformed poses and updated constraints.Further proximity-search and ICP loop closures connect the sessions for incremental joint optimization.Source: D. Map Representation and… · p. 4

E. Hyperparameter Considerations

Cross-platform robustness uses largely shared settings, while submap registration and speed-adjusted loop-closure frequency address platform variation and graph density.

E. Hyperparameter Considerations

  • MARS and NCLT use identical loop-closure ICP thresholds and FPFH/RANSAC map-merging settings, with adjustments limited to sensor-specific parameters and speed-related detection and keyframe thresholds.The system reports that extensive tuning is not required for cross-platform robustness.Source: E. Hyperparameter Considerations · pp. 4–5
  • Submap-based alignment aggregates neighboring keyframes, rejects unreliable matches by ICP error threshold, and uses voxel-downsampled point clouds with shared neighborhood radii and voxel sizes across datasets.Map merging uses submap-to-submap registration; the current submap can include up to n+1 keyframes, while the historical one can include up to 2n+1.Source: E. Hyperparameter Considerations · p. 5
  • Loop-closure detection frequency is adjusted for platform speed so keyframes remain spatially separated and chained closures do not create locally dense constraints that dominate odometry.The stated condition is v f_lc > r_lc, keeping the effective BFS search range comparable to the direct closure radius.Source: E. Hyperparameter Considerations · p. 5

IV. EXPERIMENTS

Experiments assess accuracy, consistency, alignment, and efficiency on two large-scale multi-session datasets, comparing Chain-SLAM with established baselines and an unchained ablation.

IV. EXPERIMENTS

  • The evaluation compares Chain-SLAM with Fast-LIO2 and LAMM, and isolates chained loop closure by comparing otherwise identical system settings with and without it.The datasets cover high-speed autonomous driving and moderate-speed outdoor robot scenarios.Source: §IV · p. 5
  • Trajectory accuracy uses full-trajectory APE, geometric consistency uses inter-session accumulated-point-cloud registration, and viewpoint alignment uses regional 3D reconstruction quality.The evaluation also measures runtime and memory usage.Source: §IV · p. 5
  • All experiments use the same platform: an Intel i9-12900K CPU and 64 GB DDR5 5600 MHz RAM.Source: §IV · p. 5

A. Datasets

The evaluation uses MARS and NCLT, which differ in platform, sensing density, and motion profile. Trajectory visualizations contrast the baselines and chained versus unchained Chain-SLAM.

A. Datasets

  • Both MARS and NCLT contain repeated traversals over extended durations, and each dataset comprises five sessions.The datasets enable evaluation of multi-session alignment and long-horizon geometric consistency.Source: A. Datasets · p. 5
  • Chained Chain-SLAM shows the best alignment with minimal distortion on MARS and NCLT, while unchained results show slight distortion along long roads with limited revisits.Fast-LIO2 has revisit misalignments, especially in large high-speed scenes; LAMM's global optimization introduces distortions and misalignments.Source: Figure 4 · p. 6
  • MARS is high-speed urban driving data with dense LiDAR and frequent revisits, whereas NCLT is outdoor mobile-robot data with sparser observations and more agile motion.These differing sensing and motion conditions provide complementary multi-session SLAM evaluation settings.Source: A. Datasets · p. 6

B. Trajectory Accuracy: Absolute Pose Error

Chain-SLAM reduces trajectory errors on MARS and NCLT, with chained loop closure further improving alignment. Fast-LIO2’s rotation results vary with evaluation setup and sensor conditions.

B. Trajectory Accuracy: Absolute Pose Error

  • On MARS, Chain-SLAM achieves 5–10× lower translation errors and 1.5–3× lower rotation errors than the baselines.On NCLT, it also achieves lower translation RMSE than the baselines.Source: B. Trajectory Accuracy: Absolute Pose… · p. 6
  • Chained loop closure further reduces translation and rotation RMSE over the ablation across both datasets.The reduction supports improved trajectory alignment and global geometric consistency.Source: B. Trajectory Accuracy: Absolute Pose… · p. 6
  • Fast-LIO2 has the lowest rotation RMSE on NCLT, but its per-session evaluation measures single-session accuracy rather than global alignment.On faster MARS, with a 10 Hz IMU, Fast-LIO2 has 1.5× higher rotation error than Chain-SLAM; NCLT uses a 100 Hz IMU and slower motion.Source: B. Trajectory Accuracy: Absolute Pose… · p. 6

C. Geometric Consistency: Point Cloud Registration

Point-cloud registration and 3D reconstruction assess geometric consistency and viewpoint alignment across revisited streets. Chain-SLAM, especially its chained variant, reports stronger results on both evaluations.

C. Geometric Consistency: Point Cloud Registration

  • The registration evaluation uses four MARS streets with frequent revisits from different directions and compares accumulated scans against ground-truth reference point clouds.Registration RMSE reflects inter- and intra-session geometric consistency; RANSAC is followed by ICP refinement with identical parameters and fixed random seeds across methods.Source: C. Geometric Consistency: Point Cloud… · pp. 6–7
  • Chain-SLAM reduces RANSAC RMSE by 20% over Fast-LIO2 and 40% over LAMM; the chained variant has the lowest ICP-refined errors on most streets.The chained variant preserves local geometric consistency while improving global alignment.Source: Table III · p. 7, C. Geometric Consistency: Point Cloud… · p. 7, Figure 6 · p. 7

D. Viewpoint Alignment: 3D Reconstruction

  • Chain-SLAM achieves the best reconstruction quality among compared approaches, with the chained variant strongest overall.The evaluation uses test-set 3DGS PSNR and loss on four MARS streets; higher test PSNR and lower loss indicate improved viewpoint alignment and trajectory consistency.Source: Table IV · p. 7, D. Viewpoint Alignment: 3D… · p. 7

E. Computational Efficiency

Chain-SLAM runs in real time while jointly optimizing maps from previous sessions and a current trajectory. The conclusion reports improved multi-session consistency and accuracy, with GNSS dependence remaining a scope boundary.

E. Computational Efficiency

  • The system reaches approximately 2 Hz on 12-hour MARS data and 13 Hz on 8-hour NCLT data while preserving real-time performance.Keyframe management and global optimization run asynchronously in the backend without affecting odometry propagation; memory use is slightly lower than LAMM while achieving better alignment.Source: E. Computational Efficiency · pp. 7–8

V. CONCLUSIONS

  • Across large-scale driving and mobile-robot datasets, Chain-SLAM improves trajectory accuracy, geometric consistency, and viewpoint alignment with minimal cross-platform hyperparameter tuning.The conclusion also reports real-time performance across extended multi-session deployments.Source: §V · p. 8
  • Map merging currently relies on GNSS to propose cross-session correspondences, limiting the stated approach to settings where that cue is available.The authors note LiDAR-based place recognition as a possible replacement for GNSS-free merging.Source: §V · p. 8
Loading 2609.12221v1…