PaperlayerView exact revision on arXiv

Source-linked AI summary

Degeneracy-Orthogonal Geometric Constraints for LiDAR SLAM

Minseo Kim, Yina Kim, Jinhwa Hwang, Alex Junho Lee

arXiv:2609.36753v1cs.RO

TL;DR

Axially uniform corridors leave LiDAR odometry weakly constrained along the travel direction, while repetitive structures make revisited landmarks difficult to identify. DeCOD encodes and matches cross-sectional boundary geometry, then uses verified landmarks as pose-graph constraints. Benchmark and field results report reduced trajectory error, with correction limited when reliable landmarks are scarce.

  • Problem

    In axially uniform corridors, longitudinal translation is weakly constrained, and repetitive layouts make re-identifying previously visited structures difficult.

    Source: §I · p. 1

  • Method

    DeCOD describes signed normal offsets around cross-sectional boundaries, handles association perturbations, and uses verified landmark pairs to constrain cross-section position and corridor-axis alignment.

    Source: §I · p. 1, A. Descriptor Extraction · p. 3, C. Verification and Optimization · p. 4

  • Results

    Across tunnel benchmarks and field pipe inspections, DeCOD reduces false associations and trajectory error; on Shield S8-Beta, position RMSE fell from 61.597 m to 0.768 m.

    Source: C. Trajectory Correction Performance… · p. 6, §V · p. 8

  • Takeaways & Limitations

    Verified cross-sectional landmarks provide position and axis-alignment constraints that reduce accumulated longitudinal drift while leaving rotation about the common axis unconstrained.

    Source: §V · p. 8

    On Shield S9, RMSE fell from 94.902 m to 24.819 m, but few loop closures left substantial trajectory error under the conservative acceptance policy.

    Source: C. Trajectory Correction Performance… · p. 7

Abstract

from arXiv · show

Autonomous robot navigation relies on simultaneous localization and mapping (SLAM) to estimate motion and maintain an accurate pose within an environment. However, in axially uniform corridors such as long tunnels and pipelines, LiDAR odometry is fundamentally limited by unconstrained drift along the feature-weak travel direction. This structural degeneracy cannot be resolved by local scan matching alone. To address this challenge, we propose the Degeneracy-orthogonal Contour Offset Descriptor (DeCOD), a structure-aligned geometric descriptor for cross-sectional landmarks. Cross-sectional boundaries, such as pipe joints and structural rings, provide metric constraints along this degenerate axis, but distinguishing individual landmarks requires capturing subtle surface variations across nearly identical profiles. The descriptor parameterizes signed normal deviation from estimated boundary contours, and matching explicitly resolves heading ambiguity and decouples first-order contour errors by distortion estimation. Matched landmarks yield geometric factors that enforce agreement in cross-section position and corridor axis alignment during pose-graph optimization, correcting longitudinal drift while leaving rotation about the common axis unconstrained. On a public benchmark and in field experiments, DeCOD achieves robust landmark retrieval over standard 3D descriptors and successfully stabilizes trajectories across different odometry frontends, reliably constraining longitudinal drift under geometric degeneracy.

I. INTRODUCTION

Axially uniform corridors leave longitudinal motion weakly constrained, while repetitive geometry makes revisited landmarks difficult to distinguish. DeCOD describes cross-sectional boundaries to support landmark association and drift correction.

I. INTRODUCTION

  • Repeated corridor surfaces leave translation along the travel axis weakly constrained, allowing position errors to accumulate despite locally consistent alignment.Standard place-recognition methods also struggle with the structural ambiguity of repetitive layouts.Source: §I · p. 1
  • DeCOD encodes signed normal deviations around cross-sectional boundary contours as a structure-aligned descriptor for degenerate environments.The descriptor is intended to capture surface variation that can distinguish landmarks with similar profiles.Source: §I · p. 1
  • Its association formulation addresses contour-estimation errors, partial occlusion, and reversed traversals, using geometric verification to improve precision.Source: §I · p. 1
  • The paper validates landmark association and drift reduction on a public benchmark and in field pipe experiments.Source: §I · p. 1

A. Place Recognition under Degeneracy

Repeated corridor profiles make global place recognition ambiguous, and axial degeneracy leaves longitudinal position weakly constrained. Associating discrete transverse landmarks offers localized references, but reliable matching must preserve subtle geometry under inspection perturbations.

A. Place Recognition under Degeneracy

  • Global descriptors can produce near-identical signatures at distinct corridor locations, causing perceptual aliasing and false loop closures.Discrete transverse landmarks provide localized references for longitudinal position when their geometry can be extracted.Source: A. Place Recognition under Degeneracy · p. 2

B. Degeneracy in LiDAR Odometry

  • Axial wall uniformity leaves travel-axis registration unconstrained, while restricting unreliable updates or using external beacons does not provide practical autonomous metric constraints in extended tunnels.Cylindrical radial constraints do not determine axial position; prior joint counting assumes uniform pipe lengths and known manhole locations.Source: B. Degeneracy in LiDAR Odometry · p. 2

C. Degeneracy-Aware Landmarks and Matching

  • Existing annular and contour representations motivate unwrapping, but coarse codes and histograms can discard subtle variations needed to distinguish similar landmarks under occlusion, contour error, and reversed traversal.Source: C. Degeneracy-Aware Landmarks and… · p. 2

III. DRIFT CORRECTION UNDER DEGENERACY

DeCOD addresses longitudinal drift by associating transverse boundaries as landmarks and adding their geometric constraints to pose-graph optimization. Its descriptors encode fine surface variation around detected boundaries for matching across revisits.

III. DRIFT CORRECTION UNDER DEGENERACY

  • Local scan matching cannot measure longitudinal displacement along the geometrically degenerate axis, so pose drift accumulates.DeCOD models detected cross-sectional boundaries as transverse landmark planes and uses their associations for drift correction.Source: §III · pp. 2–3
  • The pipeline detects structural boundaries, encodes signed normal offsets, compensates for contour errors, matches patches, and verifies associations before adding position and axis-alignment constraints.Source: §III · p. 3, Figure 3 · p. 3

A. Descriptor Extraction

  • Descriptor extraction estimates a corridor axis and reference contour; the passage geometry determines whether translation alone or translation plus axial rotation is degenerate.A transverse basis fixes the circumferential origin, and Difference of Gaussians with spatial non-maximum suppression selects candidate landmark origins.Source: A. Descriptor Extraction · p. 3
  • Each descriptor discretizes an axial–circumferential chart and stores mean signed normal offsets, preserving fine surface variations between similar cross-sectional profiles.The chart uses axial coordinate z and normalized circumferential arclength p around the fitted contour.Source: A. Descriptor Extraction · p. 3

B. Invariant Matching

Because accumulated drift prevents spatially bounded candidate search, DeCOD compares descriptor geometry while addressing heading reversal, incomplete observations, and contour-estimation distortions. It then smooths local patches for matching and subsequent geometric verification.

B. Invariant Matching

  • Descriptor similarity is used for candidate retrieval because accumulated longitudinal drift prevents spatial search windows, and raw matrices are unreliable under reversed travel, partial coverage, and boundary misalignment.Source: B. Invariant Matching · pp. 3–4
  • Candidates are evaluated under both 0° and 180° relative headings because reverse traversal reverses axial and circumferential order while leaving signed surface offsets invariant.Source: B. Invariant Matching · p. 4
  • Descriptors track observed cells and reject pairs with insufficient overlap so unobserved regions and noisy fragments do not corrupt matching.Source: B. Invariant Matching · p. 4
  • Least-squares fitting removes circumferential and axial biases plus tilt terms before local patches are compared with Gaussian-weighted differences.Mutual nearest-neighbor matches proceed to geometric verification.Source: B. Invariant Matching · p. 4

C. Verification and Optimization

DeCOD accepts landmark associations only after geometric-consensus verification, then uses verified pairs in trajectory optimization. Its factors constrain cross-section positions and corridor axes while leaving rotation about the common axis unconstrained.

C. Verification and Optimization

  • A candidate is accepted only when it uniquely ranks highest by geometric consensus and meets the minimum inlier-count threshold.The consensus score is 2Kij/(Ni + Nj), where Kij is the number of geometric inliers and Ni and Nj are extracted keypoint counts.Source: C. Verification and Optimization · p. 4
  • Errors in the gravity reference used for contour alignment can introduce inaccurate rotational estimates.The reference can determine rotation about the corridor axis.Source: C. Verification and Optimization · p. 4
  • Verified landmark pairs contribute position and corridor-axis alignment constraints, while the optimization preserves relative motion from the initial odometry trajectory and leaves common-axis rotation unconstrained.The objective combines odometry and landmark-factor terms under a robust Huber loss.Source: C. Verification and Optimization · p. 4
  • The pipeline proceeds from landmark detection and invariant patch matching through factor-graph trajectory optimization.Algorithm 1 retains the odometry trajectory if no verified landmark pairs are found.Source: C. Verification and Optimization · p. 4

IV. EXPERIMENTS AND RESULTS

The experiments evaluate DeCOD on the public GEODE benchmark and a custom underground pipe dataset using multiple LiDAR-inertial frontends. Figure 4 presents GEODE retrieval and trajectory-error evaluations.

IV. EXPERIMENTS AND RESULTS

  • The evaluation uses the public GEODE benchmark and a custom underground pipe dataset across multiple LiDAR-inertial frontends.The experiments examine landmark retrieval, axial localization, longitudinal-drift correction, component contributions, and computational efficiency.Source: §IV · pp. 4–5
  • Figure 4(a) shows Shield precision–recall curves at 2.4 m axial tolerance, while Figure 4(b) shows logarithmic position error on Shield S10-Beta.Panel (b) compares BIEVR-LIO and COIN-LIO before and after DeCOD correction.Source: Figure 4 · p. 5

A. Evaluation Setup

The setup models return-traversal landmark matching as an opportunity for loop closure: verified boundary associations connect the traversals through position and axis-alignment constraints. It evaluates datasets spanning different tunnel and pipe conditions and specifies local-submap descriptor construction.

A. Evaluation Setup

  • A landmark database is built during the outbound traversal and queried on return, with verified boundary associations connecting the traversals through pose-graph constraints.The constraints combine position and axis alignment and are optimized jointly with odometry factors to reduce accumulated longitudinal drift.Source: A. Evaluation Setup · p. 5

1) Datasets:

  • GEODE evaluation uses ten Shield sequences and seven Tunneling recordings, while the custom dataset has seven pipe-inspection sequences spanning tens of meters.The pipes are underground polyethylene and concrete, 400–800 mm in diameter; GEODE sequences span hundreds of meters and use Velodyne, Ouster, and Livox Avia LiDARs with synchronized IMUs.Source: 1) Datasets · p. 5

2) Descriptor Extraction:

  • Local point clouds are formed every 8 seconds from 20 motion-compensated LiDAR scans, voxelized at 5.5 cm, and used to estimate the corridor axis by PCA.Scan-wise correction reduces residual distortion before descriptor construction; axial DoG filtering and nonmaximum suppression detect boundary candidates.Source: 2) Descriptor Extraction · p. 5
  • Validated landmarks require an elongated corridor and at least 90% of projected points within 0.18 m of the reference contour.Descriptors use a 16 × 256 grid spanning ±2.4 m axially, with 256 normalized arclength bins; gravity alignment and a masked bottom ±30° sector address rotation and ground returns.Source: 2) Descriptor Extraction · p. 5

3) Matching and Optimization:

Geometric verification markedly increases DeCOD precision in the reported GEODE retrieval results, including under limited field of view. The evaluation tests verified matching and trajectory-related metrics using separated outbound-database and return-query submaps.

3) Matching and Optimization:

  • Mutual nearest-neighbor patch matches undergo geometric verification, and trajectory correction uses verified position and axis-alignment factors with relative odometry.For correction, the acceptance threshold is evaluated at K ≥9 and K ≥15; the trajectory evaluation uses at least nine inliers.Source: 3) Matching and Optimization · p. 5

4) Evaluation Protocol and Metrics:

  • Outbound database and return-query submaps use identical local point clouds, with eligible pairs separated by at least 30 seconds and sharing no scans.The protocol removes submaps without ground-truth support before candidate selection.Source: 4) Evaluation Protocol and Metrics · p. 5
  • Recall@1 and Precision@1 are evaluated against continuous ground-truth trajectories, with axial tolerances of 0.3 m and 2.4 m.Full trajectories are assessed by GEODE position RMSE and custom-pipe endpoint error; return-traversal position error over time illustrates drift reduction.Source: 4) Evaluation Protocol and Metrics · pp. 5–6

B. Place Recognition Performance in GEODE

  • Verification can improve recall by reranking candidates, because it affects both candidate selection and acceptance.The comparison evaluates DeCOD with and without geometric verification.Source: B. Place Recognition Performance in… · p. 6
  • On GEODE Shield 1–10, the reported DeCOD place-recognition results use geometric verification with K ≥15.Source: Table I · p. 6, B. Place Recognition Performance in… · p. 6
  • At 2.4 m on Tunneling, verification raised precision from 52.2% to 91.7% while accepted queries fell from 23 to 12; precision at 0.3 m was 75.0%.The Gamma Livox sensor has a limited field of view; the passage reports that fewer accepted queries substantially reduced false associations.Source: ACCEPTED QUERIES AS COUNTS. THE… · p. 6

C. Trajectory Correction Performance in GEODE

DeCOD’s verified geometric constraints reduced trajectory error in Shield sequences and preserved odometry accuracy across the Tunneling 5 sensor variants. Correction was limited when severe drift coincided with too few reliable loop closures.

C. Trajectory Correction Performance in GEODE

  • Trajectory evaluation used BIEVR-LIO base odometry, with COIN-LIO additionally tested on Shield S10-Beta; DeCOD factors required at least nine inliers.The comparison used verified position and axis-alignment constraints; STD was excluded because of near-zero axial localization performance.Source: C. Trajectory Correction Performance… · p. 6
  • On Shield S8-Beta and S10-Beta, position RMSE fell from 61.597 m to 0.768 m and from 132.817 m to 4.206 m, respectively.Table III reports trajectory position RMSE in meters for the GEODE benchmark with BIEVR-LIO base odometry.Source: C. Trajectory Correction Performance… · p. 6, Table III · p. 7
  • Across Tunneling 5’s Alpha, Beta, and limited-forward-view Gamma sensors, DeCOD preserved odometry accuracy by rejecting false candidates.When no proposed match met the verification threshold, no loop was accepted.Source: C. Trajectory Correction Performance… · p. 6
  • On Shield S9, position RMSE decreased from 94.902 m to 24.819 m, but too few loop closures left substantial error.The conservative acceptance policy limits correction when severe odometric drift coincides with insufficient reliable landmarks.Source: C. Trajectory Correction Performance… · p. 7

D. Experiment on Custom Dataset

Field trials in hume and polyethylene pipes showed that verified pipe-joint constraints reduced endpoint error across the recorded sequences. On hume 04, outbound and return scans aligned more closely after correction.

D. Experiment on Custom Dataset

  • The trials used a wheeled inspection robot on 400–800 mm hume concrete and ribbed polyethylene pipes, with outbound and return traversals.Spatial point sampling was scaled down tenfold while the descriptor grid remained 16×256; endpoint error was measured in the launch manhole frame.Source: D. Experiment on Custom Dataset · p. 7
  • Verified pipe-joint constraints reduced endpoint errors across all pipe sequences, reaching 0.022 m on hume 04 and a reduction of up to 99.9% on hume 01.On pe 03, endpoint error fell by 94.4%; hume trials reached sub-meter endpoint error in three of four trials.Source: D. Experiment on Custom Dataset · p. 7, Table IV · p. 7

E. Design and Performance Analysis

Design analyses show that contour-offset magnitudes, geometric verification, and the 5-DoF factor choice matter to DeCOD’s performance. The conclusion reports reduced false associations and trajectory error, while noting that the current formulation addresses a single dominant weak axis.

E. Design and Performance Analysis

  • On 80 Shield queries, scale-only correction reduced Recall@1 from 16.3% to 12.5%, while retaining only offset signs reduced it to 10.0%.The five-variable model compensates for contour-centering and sensor-tilt errors; the separate sign-only comparison indicates offset magnitudes also contribute.Source: E. Design and Performance Analysis · p. 7
  • At 2.4 m axial tolerance and K ≥9, geometric verification increased Shield precision from 11.7% to 74.4% and Recall@1 from 11.5% to 41.8%, while accepted queries decreased from 205 to 117.Verification improves candidate selection through geometric consensus among local patches while rejecting proposals with insufficient inliers.Source: E. Design and Performance Analysis · p. 7
  • Using Pose3 factors increased RMSE from 0.768 to 6.403 m on S8-Beta and from 4.206 to 9.708 m on S10-Beta, supporting the 5-DoF formulation.The 5-DoF factors avoid imposing a potentially erroneous rotation estimate about the corridor axis.Source: E. Design and Performance Analysis · pp. 7–8
  • Across 987 candidate pairs, descriptor extraction, matching, and verification averaged 2.15 ms per pair; a 20-candidate query took 69.7 ms versus 321.6 ms for HOTFLoc++.The paper states that sub-100 ms query latency supports periodic background evaluation on accumulated submaps.Source: E. Design and Performance Analysis · p. 8

V. CONCLUSION

  • Across tunnel benchmarks and field pipe inspections, DeCOD’s normal-offset descriptors, local patch matching, and geometric verification reduced false associations and trajectory error.Its factors constrain cross-section position and corridor-axis alignment while leaving rotation about the common axis unconstrained.Source: §V · p. 8
  • The current formulation addresses a single dominant weak axis along uniform passages; extending directional constraints to degeneracy in multiple degrees of freedom remains future work.Without external positioning references, subterranean SLAM relies on relative loop constraints only.Source: §V · p. 8
Loading 2609.36753v1…