Intensity Scan Context: Coding Intensity and Geometry Relations for Loop Closure Detection

Han Wang, Chen Wang, Lihua Xie

TL;DR

Most LiDAR-based SLAM systems ignore the "intensity" (reflectance) value of laser returns, relying solely on 3D geometry. This paper introduces Intensity Scan Context (ISC), a global descriptor that integrates both geometry and intensity to achieve highly accurate loop closure detection, even in challenging "reverse visit" scenarios, at a blistering speed of 1.2 ms per query.

What Problem Does This Paper Address?

In Simultaneous Localization and Mapping (SLAM), "loop closure detection" is the ability of a robot to recognize a place it has seen before. This is crucial for correcting the "drift" (accumulated error) in the robot's estimated trajectory.

While cameras struggle with lighting changes and weather, LiDAR is robust. However, most LiDAR methods only look at the shape of the environment. The authors argue that geometry alone can be ambiguous—two different hallways or forest roads might look identical. They tap into the intensity channel, which records how reflective a surface is (e.g., a metal sign vs. a concrete wall). This extra "color-like" information makes places much more distinguishable.

Methods

The authors propose a novel global descriptor and a two-stage matching process:

  1. Intensity Scan Context (ISC): The 3D point cloud is projected onto a 2D polar grid (rings and sectors). For each cell in this grid, the descriptor stores the maximum intensity value of the points within it. This creates a 2D matrix that captures both the spatial structure and the material properties of the scene.
  2. Two-Stage Hierarchical Retrieval:
    • Stage 1: Fast Geometry Re-identification. The intensity matrix is simplified into a binary matrix (is there a point here or not?). The system uses fast bitwise XOR operations to quickly filter the database for candidates. This stage also identifies the potential rotation (yaw) between the current scan and the database scan.
  3. Stage 2: Intensity Structure Matching. For the top candidates, the system performs a more detailed comparison using cosine distance on the actual intensity values.
  4. Consistency Verification: To prevent "false positives," the system checks for temporal consistency (do the previous few scans also match?) and geometrical consistency using Iterative Closest Point (ICP) to ensure the 3D alignment is physically possible.

Datasets

The researchers evaluated their approach using:

  • KITTI Dataset: A standard benchmark for autonomous driving. They specifically tested on sequences 00, 05 (mostly forward visits), and 02 (which includes challenging "reverse visits" where the robot returns from the opposite direction).
  • Local Warehouse Robot: Data collected from an autonomous guided vehicle (AGV) equipped with a Velodyne VLP-16, navigating a real-world warehouse environment.

Key Insights and Ideas

  • Intensity as a "Signature": Intensity isn't just noise; it reveals the "reflectance structure" of an environment. Metal plates return high values, while concrete returns low values. This provides a unique "fingerprint" for locations that geometry alone might miss.
  • The Power of Binary Operations: By converting the complex 3D data into a binary 2D matrix for the first stage of matching, the authors achieved massive speedups. Bitwise XOR is significantly faster than floating-point math.
  • Solving the "Reverse Visit" Problem: Many descriptors fail if you enter a room from the opposite door. Because ISC uses a polar representation, a change in heading simply results in a "column shift" in the descriptor matrix, which is easily handled during the matching phase.

Key Results

  • Superior Accuracy: ISC achieved higher recall and precision than existing geometric-only methods like the original Scan Context or GLAROT3D. On KITTI sequence 02 (the hardest one), ISC achieved 98% precision and 91% recall, significantly outperforming vision-based methods (DBoW2) which often fail during reverse visits.
  • Efficiency: The entire query process takes only 1.2 ms on a standard desktop computer, making it well-suited for real-time robotics applications.
  • Robustness: The system successfully corrected trajectory drifts in a warehouse robot where vision-based SLAM failed due to the lack of visual features and viewpoint changes.

Limitations and Future Work

The authors note that their method still struggles in non-residential areas (like rural roads lined with trees). In these environments, both the geometry and the intensity readings are highly similar and lack distinct landmarks, leading to potential failures.

Future work could involve more sophisticated intensity calibration to further reduce noise caused by distance and incidence angles, or exploring how to better handle highly dynamic environments where objects move frequently.

Summary updated

Abstract

Loop closure detection is an essential and challenging problem in simultaneous localization and mapping (SLAM). It is often tackled with light detection and ranging (LiDAR) sensor due to its view-point and illumination invariant properties. Existing works on 3D loop closure detection often leverage on matching of local or global geometrical-only descriptors which discard intensity reading. In this paper we explore the intensity property from LiDAR scan and show that it can be effective for place recognition. We propose a novel global descriptor, intensity scan context (ISC), that explores both geometry and intensity characteristics. To improve the efficiency for loop closure detection, an efficient two-stage hierarchical re-identification process is proposed, including binary-operation based fast geometric relation retrieval and intensity structure re-identification. Thorough experiments including both local experiment and public datasets test have been conducted to evaluate the performance of the proposed method. Our method achieves better recall rate and recall precision than existing geometric-only methods.
BibTeX
@inproceedings{wang2020intensity,
  author = {Han Wang and Chen Wang and Lihua Xie},
  title = {{Intensity Scan Context: Coding Intensity and Geometry Relations for Loop Closure Detection}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2020},
  url = {https://openreview.net/forum?id=CenQaGEsDB},
}

Loading related content...

Figure