Abstract
This paper presents a method to detect reflection of 3D light detection and ranging (Lidar) scans and uses it to classify the points and also map objects outside the line of sight. Our software uses several approaches to analyze the point cloud, including intensity peak detection, dual return detection, plane fitting, and finding the boundaries. These approaches can classify the point cloud and detect the reflection in it. By mirroring the reflection points on the detected window pane and adding classification labels on the points, we can improve the map quality in a Simultaneous Localization and Mapping (SLAM) framework. Experiments using real scan data and ground truth data showcase the effectiveness of our method.
Introduction: The "Glass Problem" in Robotics
For autonomous robots, glass is a significant obstacle. Most Lidar (Light Detection and Ranging) sensors rely on the time-of-flight of laser pulses to measure distance. When a laser hits a reflective surface like a window or mirror, the light may pass through, reflect off the surface, or do both. This leads to two major issues in Simultaneous Localization and Mapping (SLAM): "ghost obstacles," where reflections create phantom rooms in the map, and "invisible walls," where the robot fails to detect the glass itself, leading to potential collisions.
While previous research has addressed 2D Lidar reflections, 3D Lidar presents much higher data volume and more complex geometry. This research aims to bridge this gap by detecting reflective planes and utilizing the "erroneous" data to improve the map rather than simply discarding it.
Methodology
The researchers developed a pipeline that processes every 3D Lidar scan through several specialized filters:
- Intensity Peak Detection: Glass reflects most strongly when the laser beam hits it perpendicularly. The algorithm traverses Lidar "rings" to find local intensity spikes, which often indicate the center of a glass pane.
- Dual Return Analysis: Utilizing the Velodyne HDL-32E’s "dual return" mode, the system records both the strongest and the last pulse returned. If these returns differ significantly in distance, the system flags a potential reflection.
- Plane Fitting and Boundary Finding: Using RANSAC (Random Sample Consensus), the system fits a geometric plane to detected glass points and searches for physical boundaries (frames or walls) to define the window's extent.
- Reflection Back Mirroring: Instead of discarding reflected "ghost" points, the authors use a Householder transformation to mirror those points across the detected glass plane. This effectively "unfolds" the reflection, placing points back where the real objects are located in the physical world.
Experimental Evaluation
The authors collected real-world data using the MARS Jackal Mapper, a mobile robot platform equipped with two Velodyne HDL-32E Lidars. The testing environment was the "STAR Center" at ShanghaiTech University, featuring corridors with floor-to-ceiling windows and glass railings. To validate results, a professional FARO 3D Scanner was used to create a high-precision ground truth map.
Key Results
The proposed method was integrated into the HDL Graph SLAM framework with the following results:
- Classification Accuracy: The system achieved 99.9% accuracy in identifying glass points and 92.8% accuracy in identifying reflected points within a 0.15m error margin.
- Map Quality: The "Classified SLAM" significantly reduced erroneous reflection points compared to raw Lidar data.
- Mapping the Unseen: The system successfully mapped the back sides of pillars and ceilings that were never in the robot's direct line of sight, purely by utilizing reflections off windows.
Discussion and Conclusion
The core insight of this work is that reflections are data, not noise. By treating a window like a mirror in a ray-tracing simulation, a robot can "see" around corners. However, the method is sensitive to plane estimation; tiny errors in estimating a glass plane's tilt can lead to large positional errors for mirrored points. Future work will focus on integrating reflection detection more tightly into the SLAM optimization loop to use glass planes as landmarks for improved localization.
Summary updated