Abstract
In LiDAR sensing, glass, mirrors, and other reflective materials often cause inconsistent data readings because laser beams may report the distance of the glass, the distance of an object behind the glass, or the distance to a reflected object. These inconsistencies pose significant challenges for robotics and 3D reconstruction, particularly regarding localization, mapping, and navigation. While existing methods utilize dual-return LiDARs to detect glass planes in single scans, this work proposes a global, optimized map of reflective planes to classify LiDAR readings across an entire trajectory. Our experiments demonstrate that this global approach provides superior classification accuracy compared to single-scan methods. The code and data for this work are available as open source.
Introduction
For mobile robots, reflective surfaces like glass and mirrors are "sensory nightmares." Standard LiDAR (Light Detection and Ranging) sensors operate by emitting laser pulses and measuring the time-of-flight for the return signal. However, when a beam encounters a reflective surface, the interaction is complex. Three primary scenarios occur: the beam reflects off the surface (reporting the glass's position), passes through the glass (reporting an object behind it), or reflects off the surface to hit a distant object elsewhere, creating a "ghost" or reflection point.
These phenomena lead to "ghost" obstacles in maps, localization failures, and navigation hazards. While previous research attempted to filter these artifacts on a scan-by-scan basis, such methods often lack the temporal consistency required for high-fidelity 3D reconstruction. This paper introduces a framework to treat reflective surfaces as persistent structural landmarks, building a globally consistent map to improve both mapping and localization.
Methodology
The authors propose two distinct pipelines to handle reflections: Reflection Detection via Plane Optimization and Reflection SLAM.
Reflection Detection via Plane Optimization
This method assumes that external pose data (e.g., from a standard SLAM algorithm) is available. The system detects potential reflective points in individual scans using two primary techniques:
- Intensity Peak Analysis: Glass reflects most strongly when the laser hits it perpendicularly.
- Dual-Return Analysis: Utilizing sensors that report two distances for a single beam allows the system to distinguish between the glass surface and objects behind it.
Once detected, the system fits planes to these points, transforms them into a global coordinate system, and optimizes their parameters and boundaries over time. To handle complex architectural features, the system utilizes 3D Convex Hulls rather than simple rectangular assumptions, providing a more robust definition of plane boundaries.
Reflection SLAM
For more autonomous operation, the system performs Simultaneous Localization and Mapping (SLAM) by extracting both "normal" planes (walls, floors) and reflective planes to serve as landmarks. Using a Gauss-Newton optimization framework, the system calculates the robot's pose by matching these planes across consecutive frames.
Reflection Back Mirroring
A standout feature of this framework is "Reflection Back Mirroring." Once a reflective plane is accurately mapped, the "ghost" points can be mathematically mirrored back across the plane. This allows the robot to "see" and map objects that are actually behind it or around a corner, effectively expanding the sensor's field of view beyond physical line-of-sight.
Experimental Setup
The research utilizes the 3DRef dataset, a large-scale benchmark specifically designed for reflection detection. The primary sensor used is the Hesai Pandar QT64, a dual-return LiDAR with an ultra-wide vertical field of view (104.2°). This wide field of view is particularly useful for capturing reflections from floors and ceilings. The dataset includes diverse indoor scenes, such as offices and corridors, with ground truth labels categorizing points as Normal, Glass, Mirror, Reflection, or Obstacle behind glass.
Results and Discussion
The "Plane Optimized" approach demonstrated significant improvements over single-scan methods and challenged specialized deep learning models.
- Indoor Precision: In experimental sequences, the proposed method achieved an Indoor Precision of 99.82%, indicating it almost never misidentified a real indoor point as a reflection.
- Reflection Removal: The method achieved a Reflection Removal Rate of up to 92.63%, successfully cleaning the majority of "ghost" points from the final map.
- Comparison with Deep Learning: While deep learning models like Cylinder3D and MinkowskiNet performed well, the geometric approach provided more consistent global results without requiring massive training sets. It was particularly effective at identifying the precise boundaries of reflective surfaces.
Limitations and Future Work
Despite its success, the system faces certain challenges. Geometric Degeneracy occurs when the environment lacks sufficient non-parallel planes (at least three are required for a 3D pose). In long, featureless corridors with only a floor and two parallel walls, the system may lose a degree of freedom. Additionally, Proximity Issues arise when a real obstacle is located very close to a reflective surface, making them difficult to distinguish.
Future work will focus on differentiating between different types of reflective surfaces (e.g., distinguishing a mirror from a glass pane) and integrating more robust point-matching algorithms to handle complex, non-planar environments.
Summary updated