PaperlayerView exact revision on arXiv

Source-linked AI summary

LiDAR Iris for Loop-Closure Detection

Ying Wang, Zezhou Sun, Cheng-Zhong Xu, Sanjay Sarma, Jian Yang, Hui Kong

arXiv:1912.03825v3cs.ROcs.CV

TL;DR

LiDAR loop-closure detection needs descriptors that are both robust and fast enough for real-world SLAM. The paper proposes LiDAR Iris, an iris-inspired binary global descriptor using LoG-Gabor filtering, thresholding, and Hamming-distance matching, with Fourier-based pose handling. Across five road-scene sequences and two protocols, LiDAR Iris reports higher loop-closure detection performance than existing point-cloud global descriptors.

  • Problem

    Existing 3D LiDAR loop-detection methods are not both robust and fast enough, while descriptor methods also struggle with descriptive power or invariance.

  • Method

    LiDAR Iris converts LiDAR bird’s-eye views into iris-like images, extracts binary feature maps with LoG-Gabor filtering and thresholding, and matches them using Hamming distance with Fourier-based pose handling.

  • Results

    Across various datasets and two protocol experiments, LiDAR Iris showed higher loop-closure detection performance than existing point-cloud global descriptors.

  • Takeaways & Limitations

    LiDAR Iris provides a global LiDAR descriptor for fast and accurate loop-closure detection with descriptor-level pose invariance in a 3D (x,y,yaw) pose space.

Abstract

from arXiv · show

In this paper, a global descriptor for a LiDAR point cloud, called LiDAR Iris, is proposed for fast and accurate loop-closure detection. A binary signature image can be obtained for each point cloud after several LoG-Gabor filtering and thresholding operations on the LiDAR-Iris image representation. Given two point clouds, their similarities can be calculated as the Hamming distance of two corresponding binary signature images extracted from the two point clouds, respectively. Our LiDAR-Iris method can achieve a pose-invariant loop-closure detection at a descriptor level with the Fourier transform of the LiDAR-Iris representation if assuming a 3D (x,y,yaw) pose space, although our method can generally be applied to a 6D pose space by re-aligning point clouds with an additional IMU sensor. Experimental results on five road-scene sequences demonstrate its excellent performance in loop-closure detection.

I. INTRODUCTION

LiDAR Iris is introduced as a global descriptor designed to make 3D LiDAR loop-closure detection faster and more accurate. It converts LiDAR point clouds into iris-inspired representations and compares binary signatures using Hamming distance.

  • Existing LiDAR loop-detection solutions are not both robust and fast enough for real-world SLAM applications.
  • LiDAR Iris is a global descriptor for LiDAR point clouds developed for fast and accurate loop-closure detection.
  • The descriptor adapts the iris-inspired Daugman Rubber Sheet Model to represent LiDAR bird’s-eye views in polar coordinates and signature images.
  • Binary signature images are generated through LoG-Gabor filtering and thresholding, and similarities between two signatures are computed with Hamming distance.

II. RELATED WORK

Prior 3D loop-detection methods face limitations in descriptor distinctiveness, rotation invariance, or training requirements. LiDAR Iris addresses these concerns by encoding richer height information, extracting binary features, and supporting rotation-invariant detection.

  • LiDAR-based loop detection is motivated by the need to correct accumulated SLAM drift by recognizing previously visited areas.
  • Point-to-point methods such as ICP operate directly on point clouds when the two clouds are already roughly aligned.
  • Keypoint-based methods use local descriptors and bag-of-words matching, but distinctive and highly repeatable 3D keypoints remain challenging.
  • Global and local descriptors respectively struggle with descriptive power and invariance, while deep-learning methods require substantial training data and generalize poorly across changing conditions.
  • Compared with Scan-Context, LiDAR Iris encodes height rather than only maximum height, extracts binary features, and performs rotation-invariant loop-closure detection.

III. LIDAR IRIS

LiDAR Iris constructs a polar image from a LiDAR bird’s-eye view by encoding height within spatial bins, then expands it into an iris-like image representation. The resulting representation captures geometric structure while supporting efficient matching under pose variation.

  • Generation of LiDAR-Iris Image Representation: The pipeline projects each point cloud into a bird’s-eye view and discretizes an 80 × 80 m sensing square into 80 radial by 360 angular bins.
  • Generation of LiDAR-Iris Image Representation: Each bin can encode height, range, reflection, or ring information; the paper uses an eight-bit binary code based on occupied height intervals.
  • Generation of LiDAR-Iris Image Representation: The bird’s-eye view is expanded into an image strip using the Daugman Rubber Sheet Model, following the iris-recognition analogy.
  • Generation of LiDAR-Iris Image Representation: The eight-bit code for each bin is converted into a decimal intensity, producing a LiDAR-Iris image whose rows and columns correspond to the discretized dimensions.
  • Generation of LiDAR-Iris Image Representation: At the same geometrical location, LiDAR-Iris images are mainly related by cyclic translation when the corresponding point clouds differ by rotation.

B. Fourier transform for a translation-invariant LiDAR Iris

Fourier-transform alignment addresses translation variation when matching LiDAR-Iris images, while the paper uses the resulting alignment to support rotation-invariant matching. The method estimates image translation from the cross-power spectrum and inverse transform.

  • Fourier-transform alignment estimates translation between two LiDAR-Iris images for matching.The method models images differing by a shift and uses their Fourier-domain relationship to recover that shift.
  • Figure 5 illustrates rotation-invariant matching through Fourier-transform alignment of corresponding LiDAR-Iris images.
  • Robot translation can cause vertical image shifts and slight pixel-intensity changes after Fourier transformation.The encoding preserves the point cloud’s internal bin structure, so small translation-induced intensity changes are ignored as negligible.
  • The inverse Fourier transform produces a correlation response whose nonzero peak identifies the translation offset.The offset is selected at the argmax of the correlation response.

C. Binary feature extraction with LoG-Gabor filters

LoG-Gabor filtering decomposes LiDAR-Iris images into localized multiresolution features, which are thresholded into binary responses. Four filters provide the reported best validation accuracy at low computational cost.

  • LoG-Gabor filters extract localized features at different resolutions from LiDAR-Iris images.Compared with traditional Fourier transforms, their localized frequency data supports matching features at the same position and resolution.
  • One-dimensional LoG-Gabor filters are used to preserve real-time capability.
  • The filter’s center frequency is controlled by f0, while σ affects bandwidth; maintaining σ/f0 preserves filter shape across frequencies.
  • Four LoG-Gabor filters achieve the best loop-closure detection accuracy at low computational cost on the validation dataset.The first four filters are therefore used for the experimental results, and their convolution responses are binarized by thresholding.

IV. LOOP-CLOSURE DETECTION WITH LIDAR IRIS

LiDAR Iris performs loop-closure detection by comparing binary feature maps for the current keyframe against historical keyframes using Hamming distance and a threshold.

  • A binary LiDAR-Iris feature map is generated for each keyframe and stored in a history database during traversal.
  • Hamming distance compares the current keyframe’s binary feature map with each historical keyframe.
  • A distance below the threshold is classified as a loop-closure event.Loop closure then triggers backend optimization to correct estimated poses and maps in full SLAM.

V. EXPERIMENT

The experiments compare LiDAR Iris with three other global 3D point-cloud descriptors under the same computing setup. The supplied passages specify the comparison methods and hardware conditions, but no performance values.

  • LiDAR Iris is compared with Scan-Context, M2DP, and ESF because all extract global descriptors for 3D point clouds.
  • All experiments use the same PC with an Intel i7-8550U CPU at 1.8 GHz and 8 GB memory.

A. Dataset

Experiments use three KITTI and two campus sequences with diverse LiDAR sensors and loop directions, evaluated under real loop-closure and place re-identification protocols.

  • Datasets: Five sequences comprise three KITTI odometry datasets and two campus datasets collected with 64-channel and 16-channel LiDAR sensors.The datasets include same-direction and opposite-direction loop events.
  • Evaluation protocols: Precision and recall are evaluated under both protocols, with true positives in Protocol A additionally requiring ground-truth distance below 4m.The 4m criterion is adopted as the default according to the cited prior work.
  • Evaluation protocols: Protocol A matches each current keyframe against previous database keyframes, excluding the 30 immediately preceding frames, and thresholds the closest feature distance.A prediction is a loop closure when the closest feature distance is no larger than d_f.
  • Visualization: Affinity-matrix visualizations compare ground truth, LiDAR Iris, Scan-Context, M2DP, and ESF on KITTI 05 and the smaller campus scene.The figures also show trajectories, reconstructed maps, and corresponding loop-closure areas for the campus and KITTI data.
  • Evaluation protocols: Protocol B formulates loop detection as place re-identification by labeling keyframe pairs positive when their locations are within 4m and negative otherwise.Pairwise feature distances produce an affinity matrix that is thresholded for evaluation.

C. Performance Comparison

LiDAR Iris is compared with Scan-Context, M2DP, and ESF across five sequences using precision-recall curves and affinity matrices under two protocols; it achieves the strongest reported performance overall.

  • Experimental setup: All methods use raw point clouds without downsampling, while Scan-Context, M2DP, and ESF use their stated or available default parameters.The comparison evaluates all candidate keyframes rather than applying candidate selection parameters.
  • Precision-recall evaluation: Precision-recall curves cover KITTI 00, KITTI 05, KITTI 08, and smaller and larger campus scenes under Protocols A and B.The figure’s first and second rows correspond to Protocols A and B, respectively.
  • Affinity matrices: LiDAR Iris and Scan-Context effectively reveal loop-closure regions in exemplar affinity matrices, while M2DP and ESF perform substantially worse.Scan-Context also produces low matching values for some negative pairs, which can cause false positives.
  • Precision-recall evaluation: ESF shows the worst performance on every sequence under both protocols, relying on histograms that distinguish places mainly when visible structures differ substantially.This limitation is attributed to its histogram-based representation.
  • Protocol comparison: Under Protocol A, LiDAR Iris achieves the best performance among the four methods across all five sequences, whereas M2DP fails on opposite-direction or mixed-direction loops.The paper attributes LiDAR Iris’s advantage to discriminative binary features and Fourier-transform translation invariance.
  • Protocol comparison: Under Protocol B, the paper reports that LiDAR Iris again achieves the best performance.Protocol B evaluates many more matching pairs, with negative pairs greatly outnumbering positive pairs.

D. Computational Complexity

Computational complexity is measured for pairwise matching on KITTI 00, excluding LiDAR Iris generation and candidate retrieval; LiDAR Iris is slightly faster than Scan-Context.

  • Runtime comparison: LiDAR Iris requires about 0.0231s per frame, compared with 0.0257s for Scan-Context.These are average matching times measured in Matlab on KITTI sequence 00.
  • Measurement scope: LiDAR Iris complexity includes binary-feature extraction and matching two binary maps, but excludes LiDAR-Iris generation.Scan-Context complexity is measured by the time required to match two Scan-Context images.
  • Measurement scope: The timing compares matching every pair of frames without ring-key candidate parameters, so it is independent of candidate count.The PR comparison likewise uses all candidate keyframes.

VI. CONCLUSION

The paper concludes that LiDAR Iris represents places as binary signature images and delivers higher loop-closure detection performance across varied datasets and evaluation protocols.

  • Contribution: LiDAR Iris summarizes a LiDAR point cloud as a binary signature image produced by Gabor filtering and thresholding its image representation.The descriptor is proposed for global LiDAR point-cloud representation and loop-closure detection.
  • Conclusion: Across various datasets and two protocol experiments, LiDAR Iris shows higher loop-closure detection performance than existing point-cloud global descriptors.This is the paper’s stated comparative conclusion.
Loading 1912.03825v3…