PaperlayerView exact revision on arXiv

Source-linked AI summary

Local Descriptor for Robust Place Recognition using LiDAR Intensity

Jiadong Guo, Paulo V. K. Borges, Chanoh Park, Abel Gawel

arXiv:1811.12646v1cs.RO

TL;DR

LiDAR place recognition needs appearance information that geometry-only methods lack, especially in challenging environments. The paper combines calibrated intensity with geometry in ISHOT and adds probabilistic keypoint voting, achieving strong local-descriptor and global-localization performance while exposing sensor-calibration limits.

  • Problem

    LiDAR place recognition lacks standardized, transferable intensity information, while geometry-only methods can struggle in natural environments with cluttered vegetation and noisy surface normals.

  • Method

    The paper combines calibrated LiDAR intensity with SHOT geometry in ISHOT and uses probabilistic keypoint voting based on empirically modeled voting precision.

  • Results

    The approach consistently achieves success rates over 90%, including 100% on the gator dataset, and outperforms reference algorithms on more challenging datasets by over 10%.

  • Takeaways & Limitations

    The framework provides competitive real-time global place recognition that is robust to viewpoint changes and occlusion in challenging outdoor experiments.

  • Takeaways & Limitations

    Intensity localization across different LiDAR sensors remains difficult because sensor preprocessing is unobservable and raw sensor energy return is unavailable.

Abstract

from arXiv · show

Place recognition is a challenging problem in mobile robotics, especially in unstructured environments or under viewpoint and illumination changes. Most LiDAR-based methods rely on geometrical features to overcome such challenges, as generally scene geometry is invariant to these changes, but tend to affect camera-based solutions significantly. Compared to cameras, however, LiDARs lack the strong and descriptive appearance information that imaging can provide. To combine the benefits of geometry and appearance, we propose coupling the conventional geometric information from the LiDAR with its calibrated intensity return. This strategy extracts extremely useful information in the form of a new descriptor design, coined ISHOT, outperforming popular state-of-art geometric-only descriptors by significant margin in our local descriptor evaluation. To complete the framework, we furthermore develop a probabilistic keypoint voting place recognition algorithm, leveraging the new descriptor and yielding sublinear place recognition performance. The efficacy of our approach is validated in challenging global localization experiments in large-scale built-up and unstructured environments.

I. INTRODUCTION

The paper addresses LiDAR place recognition in challenging natural and occluded environments by combining calibrated intensity with geometry. It introduces ISHOT and probabilistic keypoint voting to improve descriptor matching and localization.

  • Natural environments challenge geometric descriptors because vegetation can clutter scenes and produce noisy surface-normal estimates.
  • A flexible localization approach using intensity-based local 3D descriptors had not previously been demonstrated.
  • ISHOT combines calibrated LiDAR intensity returns with geometric information in an intensity-enriched local 3D descriptor.
  • The proposed probabilistic keypoint voting method models voting precision and updates place-matching probabilities after each vote.
  • The method is evaluated in large-scale outdoor experiments spanning 160,000 m2.
  • The framework follows the local-descriptor place-recognition pipeline while replacing geometry-only description with intensity-aware matching and probabilistic voting.

III. INTENSITY-AUGMENTED 3D DESCRIPTOR

The descriptor pipeline preprocesses VLP-16 intensity measurements, calibrates them into estimated surface intensities, and combines the resulting texture cue with SHOT geometry. Measurements beyond 30 m are discarded because intensity degrades at long range.

  • The method calibrates VLP-16 intensity measurements and introduces ISHOT as an intensity-augmented 3D descriptor.
  • Measurements beyond 30 m are discarded before calibration because intensity returns degenerate at high range.
  • Raw VLP-16 returns are noisy, differ across beams, and become degenerate at large distances.
  • The calibration maps each beam’s discrete measured intensity to the most likely true surface intensity.
  • Calibrated intensity is rescaled to [0, 1], with original values of 100 and above mapped to 1.

B. Constructing ISHOT

ISHOT extends a generalized multi-cue histogram descriptor by combining geometric SHOT information with calibrated LiDAR intensity differences. Its intensity component uses the same spatial support regions as SHOT and is concatenated with the original descriptor.

  • ISHOT is a chain of generalized Signatures of Histograms for multiple point-wise data cues around a feature point.
  • The descriptor combines geometric SHOT with a texture cue based on calibrated intensity returns.
  • Its matching metric compares each sample inside the support region with the feature point using point-wise cue properties.
  • The intensity-difference histograms use 31 bins across each of 32 spatial support regions, while the original SHOT contributes 352 dimensions.
  • The resulting ISHOT configuration has 1344 feature dimensions.

IV. PROBABILISTIC KEYPOINT VOTING

The place-recognition method adapts keypoint voting by using ISHOT descriptors with ISS-BR detection and by modeling voting accuracy probabilistically. This addresses environment-specific keypoint densities while updating probabilities for candidate places.

  • The method combines ISHOT local descriptors with ISS-BR keypoint detection within a keypoint-voting pipeline.
  • ISS-BR improves matching performance but produces environment-specific keypoint densities.
  • The probabilistic voting approach models closest-neighbor voting accuracy to update probabilities of correct place matches.
  • The system extracts ISHOT descriptors from a calibrated local 3D LiDAR scan and matches them against a global database of descriptors.

B. Global Places Database & Localization Query

The system represents a mapped trajectory as overlapping discrete places and localizes new scans by descriptor matching followed by probabilistic voting. Votes use nearest-neighbor quality and update a normalized probability distribution over database places before geometric refinement.

  • B. Global Places Database & Localization Query: The global map is partitioned into discrete, nearby-overlapping places containing voxelized point clouds with corrected and averaged intensity values.
  • B. Global Places Database & Localization Query: A localization query processes a local LiDAR point cloud similarly to database places, then matches each resulting descriptor against descriptors from all places.
  • B. Global Places Database & Localization Query: Each vote considers the two nearest descriptor neighbors, assigning the nearest neighbor’s place as candidate ρv and recording NNDR quality τv.
  • B. Global Places Database & Localization Query: The system first extracts ISHOT descriptors from a two-rotation scan, matches them to the global database, and votes for the most likely place candidate.
  • B. Global Places Database & Localization Query: After match probability exceeds a threshold, the system performs geometric consistency refinement between the scan and database candidates.
  • B. Global Places Database & Localization Query: Each vote retrieves a precomputed probability from a lookup table and updates the probability distribution over all database places rather than selecting one place directly.
  • B. Global Places Database & Localization Query: The matching probability is normalized across places and incorporates voting quality and spatial relationships, with additional normalization for differing keypoint densities.

D. Modeling voting precision

The voting process is modeled as a mixture of a half-normal distribution for spatially plausible matches and a uniform distribution for random non-matches. Parameters are fitted separately across matching-quality ranges using ground-truth training data.

  • D. Modeling voting precision: Votes are modeled with a half-normal component centered at zero and a uniform component representing random non-matches.Distances to the ground-truth place are used to fold the normal likelihood into a half-normal distribution.
  • D. Modeling voting precision: The matching probability depends on place distance, matching score τ, and fitted parameters σ and λ.λ balances the two mixture probabilities, while σ is the normal-distribution variance.
  • D. Modeling voting precision: Parameters σ and λ are obtained by fitting the theoretical cumulative distribution function to ground-truth-matched training data for each τ range.The fitting procedure is illustrated for two τ ranges.
  • D. Modeling voting precision: Lower τ ranges indicate higher matching quality but occur less frequently in the vote distribution.The vote distribution has a uniformly distributed long tail extending to 400 m, omitted from the figure for visibility.

E. Probability update and terminate condition

The pipeline avoids exhaustive descriptor computation by processing random subsets of unprocessed keypoints and updating database-place probabilities from their votes. It terminates candidate search when a place’s probability exceeds an acceptance threshold, then verifies the candidate geometrically.

  • E. Probability update and terminate condition: Each iteration computes and matches descriptors for a randomly selected subset of unprocessed keypoints.This selective strategy reduces descriptor computation and matching in the high-dimensional feature space.
  • E. Probability update and terminate condition: Votes update a normalized probability for every database place using a precomputed table indexed by matching quality and spatial relationship.The probability acts as a confidence that a vote originated from each place.
  • E. Probability update and terminate condition: Once a place probability surpasses the acceptance threshold, the system roughly aligns the scan against that candidate using geometrically consistent local-feature transformations.Candidate verification is repeated when multiple places exceed the threshold.

V. EXPERIMENTS

Experiments evaluate the approach on QCAT outdoor LiDAR data spanning a large mapped area and multiple scan conditions. The datasets vary viewpoint and occlusion while using the same sensor setup.

  • V. EXPERIMENTS: The evaluation benchmarks ISHOT descriptors and the complete probabilistic voting pipeline on real-world datasets.Descriptor evaluation uses AUC, while the integrated system is compared with reference localization approaches.
  • V. EXPERIMENTS: The QCAT environment contains 438 mapped places covering approximately 160,000 m2.The map was created with a rotating 3D LiDAR mounted on the autonomous Gator platform.
  • V. EXPERIMENTS: The gator dataset contains 58 static scans collected at diverse map locations.These scans were generated by the mapping vehicle.
  • V. EXPERIMENTS: The pole dataset contains 41 static scans collected from heights between 0.8 m and 2.5 m to create viewpoint differences.The scans use the same sensor module independently of the mobile platform and are gravity aligned.
  • V. EXPERIMENTS: Figure 6 maps the QCAT trajectory, intensity-colored environment, and ground-truth locations for the gator, pole, and occlusion datasets.The figure also shows two example places.
  • V. EXPERIMENTS: A 120 seconds driving sequence was recorded as a calibration dataset for the LiDAR intensity sensor.The calibration sequence was collected with the Gator platform at QCAT.

B. Evaluation of local descriptors

Local descriptor evaluation compares calibrated and raw-intensity ISHOT with geometric descriptors under matched neighborhood settings. ISHOT outperforms the benchmarked descriptors, while calibration further improves descriptiveness and changes the precision–recall trade-off.

  • B. Evaluation of local descriptors: Local descriptors are evaluated on gator and pole scans using precision, recall, and AUC, with true positives within 5 m of ground truth.Descriptors are matched against nearest neighbors in the map descriptor database.
  • B. Evaluation of local descriptors: All descriptors use comparable point-cloud volumes, with a 7 m radius and 12 m height for structural descriptors.Measurements beyond 40 m are discarded and point clouds are voxel-downsampled at 0.4 m.
  • B. Evaluation of local descriptors: ISHOT significantly outperforms all benchmarked descriptors with the ISS-BR detector, demonstrating place disambiguation from intensity returns.ISHOT uses raw intensity, whereas ISHOT-C uses calibrated intensity.
  • B. Evaluation of local descriptors: 73.85% and 59.35% improvements over SHOT are reported for calibrated-intensity ISHOT.Calibration maps distinctive measurements toward fewer statistically dominant values, improving descriptor similarity for similar surfaces.
  • B. Evaluation of local descriptors: ISHOT-C improves overall recall while giving up some precision at low τ compared with ISHOT.Table II also reports feature-description and matching times for the evaluated descriptors.

C. Evaluation of place recognition

The probabilistic place recognition pipeline achieves high localization success across challenging datasets while reducing matching effort through probabilistic termination. Its ISHOT matching remains the principal runtime bottleneck in the driving evaluation.

  • Probabilistic voting: The probabilistic voting approach terminates matching much earlier than the original voting process while achieving higher precision.The method models matching probability and spatial relationships between places; industrial scans typically terminate after one batch.
  • Localization performance: Success rates exceed 90% consistently, reach 100% on the gator dataset, and exceed reference algorithms by over 10% on more challenging datasets.The evaluation compares success ratio and average and median processing time across global localization pipelines.
  • Runtime: Processing times are similar to DELIGHT but slightly faster on average, with median times often much lower because only a few difficult scans require extensive processing.Difficult scans may require extracting and matching all features or examining multiple candidates, increasing average time.
  • Runtime: On a challenging driving dataset, the system achieves an averaged global localization update rate of 0.25 Hz.The scene contains dynamic objects and vehicle-motion distortion.
  • Runtime: High-dimensional ISHOT descriptor matching is the current bottleneck of the algorithm.The bottleneck is identified in the profiled driving-dataset evaluation.

VI. DISCUSSION & LIMITATIONS

The discussion identifies calibration and cross-sensor transfer as important boundaries for LiDAR intensity, while concluding that ISHOT and probabilistic voting provide robust outdoor localization performance. The authors call for more general intensity procedures and calibration standards across LiDAR sensors.

  • Limitations: Unobservable sensor preprocessing and unavailable raw energy returns make localization between different LiDAR sensors difficult.The discussion attributes preprocessing to distinguishing retro-reflective and diffusive objects.
  • Limitations: The authors identify a clear need for general LiDAR intensity calibration standards because intensity benefits robot localization.Calibration is presented as a requirement for broader use of intensity values.
  • Contributions: ISHOT combines LiDAR geometry with calibrated intensity returns and outperforms state-of-the-art geometric descriptors in real-world local descriptor evaluations.The descriptor is presented as combining geometric and texture information from a LiDAR sensor.
  • Conclusions: The framework achieves competitive real-time global place recognition performance while remaining robust to viewpoint changes and occlusion.The evaluation uses challenging outdoor experiments.
Loading 1811.12646v1…