Loop Closure Detection in 3D Laser-Based SLAM Using Spherical Harmonics Descriptors

Zikuan Li, Mingqiang Wei, Qiaoyun Wu, Yanwen Guo, Kaiqin Xu, Jia Li, Jun WangJournal of Computer-Aided Design & Computer Graphics 2023 · DOI

Abstract

Loop closure detection is the core of 3D laser-based SLAM to realize autonomous positioning and navigation. Aiming at the problems of high complexity and low accuracy in current loop closure detection, a spherical harmonic energy with rotation invariance (SHE) and a spherical harmonic energy with the Z-axis rotation invariance (SHZE) are first proposed in this paper. Combining the advantages of SHE and SHZE via a “two-step search”, a novel loop closure detection algorithm for 3D laser-based SLAM, denoted as SH-LCD, is proposed. SH-LCD not only improves the richness of information extraction, but also reduces the computation complexity of loop closure detection, thus characterized by strong feature extraction ability and universality. Extensive loop closure detection evaluations on benchmark datasets including KITTI, NCLT, and Complex Urban, demonstrate that the detection accuracy of SH-LCD significantly outperforms the current state-of-the-art methods including Scan Context, M2DP, OverlapNet, etc. In addition, the efficiency of SH-LCD is high, and the time for operator calculation and operator matching are about 12.0 ms and 2.3 ms, respectively. This meets the real-time requirements of SLAM.

Introduction

In 3D Simultaneous Localization and Mapping (SLAM), loop closure detection is the process of recognizing a previously visited location to correct the cumulative "drift" in a robot's estimated trajectory. Current methods face a difficult trade-off: local descriptors often struggle with the sparsity and noise inherent in 3D LiDAR data, while global descriptors like Scan Context or M2DP may lose structural detail or struggle with viewpoint changes. Deep learning methods offer high accuracy but often require massive datasets and are computationally expensive for real-time deployment.

This paper addresses these challenges by using Spherical Harmonics (SH)—essentially the "Fourier Transform for a sphere"—to create descriptors that are both mathematically robust to rotation and computationally lightweight.

Methodology: The SH-LCD Pipeline

The proposed SH-LCD algorithm operates through a specialized pipeline:

  1. Spherical Projection: 3D point clouds are projected onto a spherical model. To handle the non-uniform distribution of LiDAR points, an equiangular grid is used.
  2. Spherical Harmonic Transform (SHT): The spherical signal is decomposed into frequency components.
  3. Dual Descriptors:
    • SHE (Spherical Harmonic Energy): A descriptor fully invariant to rotation in any direction, used for fast, coarse-grained candidate retrieval.
  4. SHZE (Spherical Harmonic Z-axis Energy): A descriptor invariant only to rotation around the vertical (Z) axis, preserving more structural detail for ground vehicles.
The SH-LCD algorithm flowchart illustrates the two-step search process using SHE and SHZE descriptors.

As shown in the flowchart, the system employs a Two-Step Search: it first uses SHE to quickly find the top $k$ most similar frames from history, then performs a refined search on these candidates using the more detailed SHZE descriptor.

Experimental Evaluation

The algorithm was evaluated on three diverse, large-scale public datasets: KITTI, NCLT (North Campus Long-Term), and Complex Urban.

Key Results

  • Accuracy: SH-LCD outperformed state-of-the-art methods like Scan Context, M2DP, and OverlapNet across almost all tested sequences, particularly in the challenging NCLT and Complex Urban datasets.
  • Speed: Descriptor generation takes approximately 12.0 ms, and matching takes only 2.3 ms, well within the 100ms window required for real-time 10Hz SLAM.
  • Drift Correction: When integrated into the LeGO-LOAM framework, SH-LCD successfully identified loops missed by the original system, significantly reducing the Absolute Trajectory Error (ATE).
Precision-Recall curves across various datasets demonstrate that SH-LCD (red line) consistently achieves higher accuracy than competing methods.

Conclusion

SH-LCD offers a robust and efficient solution for loop closure detection in 3D LiDAR SLAM. By leveraging the natural spherical geometry of LiDAR data and a hierarchical search strategy, it achieves state-of-the-art accuracy while maintaining the low latency required for autonomous navigation. Future work will focus on improving robustness to translational offsets in wide-open environments.

Summary updated

Abstract

<p indent="0mm">Loop closure detection is the core of 3D laser-based SLAM to realize autonomous positioning and naviga-tion. Aiming at the problems of high complexity and low accuracy in current loop closure detection, a spherical harmonic energy with rotation invariance (SHE) and a spherical harmonic energy with the <italic>Z</italic>-axis rotation invariance (SHZE) are first proposed in this paper. Combining the advantages of SHE and SHZE via a “two-step search”, a novel loop closure detection algorithm for 3D laser-based SLAM, denoted as SH-LCD, is proposed. SH-LCD not only improves the richness of information extraction, but also reduces the computation complexity of loop closure detection, thus characterized by strong feature extraction ability and universality. Extensive loop closure detection evaluations on benchmark datasets including KITTI, NCLT, and Complex Urban, demonstrate that the detection accuracy of SH-LCD significantly outperforms the current state-of-the-art methods including Scan Context, M2DP, OverlapNet, etc. In addition, the efficiency of SH-LCD is high, and the time for operator calculation and operator matching are about <sc>12.0 ms</sc> and <sc>2.3 ms,</sc> respectively. This meets the real-time requirements of SLAM.
BibTeX
@article{li2023loop,
  author = {Zikuan Li and Mingqiang Wei and Qiaoyun Wu and Yanwen Guo and Kaiqin Xu and Jia Li and Jun Wang},
  title = {{Loop Closure Detection in 3D Laser-Based SLAM Using Spherical Harmonics Descriptors}},
  journal = {Journal of Computer-Aided Design \& Computer Graphics},
  year = {2023},
  doi = {10.3724/sp.j.1089.2023.19748},
}
Figure