Abstract
LiDAR is a cornerstone of autonomous perception, providing critical 3D spatial data for obstacle detection and path planning. However, off-road environments present unique challenges for traditional geometric-based semantic segmentation due to the lack of structured features and the presence of objects with ill-defined boundaries. This paper explores the use of calibrated LiDAR intensity as a robust feature for material classification. By addressing range and incidence angle dependencies, the proposed method achieves significant improvements in segmenting challenging classes like "puddle" and "grass" on the RELLIS-3D dataset, demonstrating cross-platform compatibility between Velodyne and Ouster systems.
Introduction
Autonomous navigation in off-road settings is significantly more complex than in urban environments. While urban scenes offer structured geometry like curbs and lane markings, off-road terrain is characterized by diverse objects with "undefined boundaries." Most state-of-the-art LiDAR segmentation models rely heavily on geometric features—the shape and arrangement of points. However, geometry alone is often insufficient; for instance, a flat patch of grass and a flat puddle appear identical to a sensor considering only spatial coordinates.
LiDAR sensors also record "intensity," the strength of the returned laser pulse. While intensity correlates with surface reflectivity, it is inherently unstable, fluctuating based on distance (range) and the angle of incidence. This paper proposes a calibration pipeline to transform raw intensity into a stable material signature.
Methodology
The core of the approach is a calibration model that accounts for the physical principles of laser reflection. The received power is proportional to the surface reflectance, the cosine of the angle of incidence, and inversely proportional to the square of the range.
Intensity Calibration and the Near-Range Effect
A critical observation in this study is the "near-range effect." At distances under 6 meters, the relationship between distance and intensity is non-linear and exponential, deviating from the standard inverse-square law. The authors developed a correction model focusing on the 6m–60m range where the physics is more predictable.
To accurately determine the angle of incidence ($\alpha$), which is often noisy when calculated from surface normals, the authors employed a Fully Connected Network (FCN). This network predicts a refined $\alpha$ using surface normal vectors as input, allowing for more precise intensity normalization.
Cross-Platform Compatibility
To ensure the method's utility across different hardware, the authors addressed the variations in how manufacturers process intensity. They introduced a polynomial fitting function to translate Velodyne’s pre-processed 8-bit intensity data back into a raw format compatible with Ouster sensors, facilitating cross-platform deployment.
Experimental Results
The methodology was evaluated using the RELLIS-3D dataset, specifically sequences from an Ouster OS1-64 sensor, with cross-validation on Velodyne Ultra Puck data.
The intensity-based approach was compared against deep learning benchmarks SalsaNext and KPConv. The results highlight the power of reflectivity as a material fingerprint:
- Puddle Detection: The method achieved an Intersection over Union (IoU) of 47.83, more than doubling SalsaNext (23.20) and vastly outperforming KPConv (0.0), which failed to identify puddles entirely due to their lack of geometric features.
- Grass Segmentation: The method also outperformed benchmarks for the "grass" class with an IoU of 66.44.
Discussion and Future Work
While the results are promising, certain limitations remain. The "bush" class showed poor performance (13.65 IoU) due to intensity overlap with "trees," likely stemming from similar biological textures. Additionally, surface reflectivity is sensitive to environmental conditions; wet versus dry grass will yield different signatures.
This work serves as a preliminary analysis advocating for the inclusion of calibrated intensity as a supplementary input in semantic segmentation frameworks. Future research will focus on integrating these reflectivity features into more complex deep learning models to combine the strengths of both geometric and material-based perception.
Code available at: GitHub
Summary updated