When Geometry is not Enough: Using Reflector Markers in Lidar SLAM

Gerhard Kurz, Sebastian Scherer, Peter Biber, David Fleer

TL;DR

This paper introduces a hybrid Lidar SLAM approach that uses inexpensive retroreflective tape to provide "artificial geometry" in featureless environments like long, smooth corridors. By combining a robust marker detector with two novel fusion strategies, the authors demonstrate that adding just a few markers can turn a failing SLAM system into a rock-solid one in challenging industrial settings.

What Problem Does This Paper Address?

Standard Lidar-based SLAM (Simultaneous Localization and Mapping) relies on the geometric features of an environment—corners, pillars, and wall indentations—to determine a robot's position. However, in "structureless" environments like long, featureless hallways or large open warehouses, Lidar sensors suffer from the "aperture problem." Because every section of a smooth wall looks identical to the sensor, the SLAM algorithm can't accurately determine how far the robot has moved along the corridor, leading to massive longitudinal drift or total localization failure.

While previous solutions involved expensive, pre-surveyed reflector maps, this paper looks for a "middle ground": using unobtrusive markers to provide just enough artificial structure to help existing SLAM algorithms succeed without requiring a labor-intensive setup.

Methods

The authors propose a complete pipeline consisting of detection and fusion:

1. Robust Marker Detection

Simply thresholding for high-intensity Lidar returns isn't enough, as metal objects or safety vests can cause false positives. The authors' detector looks for a specific "semantic" pattern:

  • Intensity: Points must exceed a sensor-specific intensity threshold.
  • Geometry: The high-intensity points must form a strip of a known width (e.g., 5 cm) embedded within a larger planar surface (a wall).
  • Filtering: It discards detections at steep angles (over 80°) where intensity data becomes unreliable.

2. Fusion Approaches

The paper explores two ways to integrate these markers into the Normal Distributions Transform (NDT) scan-matching process:

  • NDT Layer Approach: The scan is split into "regular" and "marker" points. Scan matching is performed on both sets in parallel. Marker points are given significantly higher weight in the optimization, effectively "pinning" the robot's pose to the detected reflectors.
  • NDT Tracking Approach: This method explicitly tracks individual markers over time using the Hungarian algorithm. It adds a dedicated "tracking-based" penalty term to the SLAM cost function, ensuring the robot's estimated trajectory aligns with the observed movement of the markers.

Datasets

The researchers evaluated their system using a custom-built sensor cart in three real-world industrial environments: two production plants with featureless corridors and one outdoor storage area with metal parts and ramps.

They collected 58 recordings using four different Lidar sensors:

  • SICK LMS151
  • Pepperl & Fuchs R2000
  • Omron OS32C
  • Velodyne VLP-16 (used primarily as a 3D reference system to establish ground truth).

Key Insights and Ideas

  • Artificial Structure vs. Infrastructure: You don't need a perfectly surveyed map of reflectors. Just placing a few pieces of tape provides the "geometric anchor" the Lidar needs to resolve its position along a hallway.
  • Intensity is Relative: One of the most striking findings is how much intensity values vary between manufacturers (see Fig. 3 in the paper). A value of 1000 might be "bright" for one sensor and "dark" for another, necessitating sensor-specific calibration.
  • Sparsity Handling: Because markers often consist of only a few Lidar points, the authors synthesized additional points around detected markers to help the NDT algorithm converge more reliably.

Key Results

The results show a dramatic improvement in SLAM reliability:

  • Success Rates: For the Omron OS32C sensor—which struggled the most—the success rate jumped from a dismal 21.4% (without markers) to 100% using the proposed NDT Layer approach.
  • Detector Precision: The proposed geometric detector significantly outperformed simple intensity thresholding, maintaining high precision even in environments cluttered with other reflective materials like metal pallets.
  • Algorithm Preference: The "NDT Layer" approach generally outperformed the "Tracking" approach, as it handled the inherent sparsity of marker points more gracefully.

Limitations and Future Work

The system's main weakness is its reliance on marker visibility. If markers are placed too sparsely or the robot views them at extremely shallow angles, the system reverts to standard geometric SLAM and may fail again.

The authors suggest that future work could involve creating "sensor-specific recommendations" for marker placement—essentially a guide for warehouse managers on exactly where to stick the tape to ensure 100% SLAM uptime. They also suggest exploring the use of these markers for global relocalization (helping a "lost" robot find itself).

Summary updated

Abstract

Lidar-based SLAM systems perform well in a wide range of circumstances by relying on the geometry of the environment. However, even mature and reliable approaches struggle when the environment contains structureless areas such as long hallways. To allow the use of lidar-based SLAM in such environments, we propose to add reflector markers in specific locations that would otherwise be difficult. We present an algorithm to reliably detect these markers and two approaches to fuse the detected markers with geometry-based scan matching. The performance of the proposed methods is demonstrated on real-world datasets from several industrial environments.
BibTeX
@inproceedings{kurz2022when,
  author = {Gerhard Kurz and Sebastian Scherer and Peter Biber and David Fleer},
  title = {{When Geometry is not Enough: Using Reflector Markers in Lidar SLAM}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2022},
  doi = {10.1109/iros47612.2022.9981522},
}

Loading related content...

Figure