Off-Road LiDAR Intensity Based Semantic Segmentation

Kasi Viswanath, Peng Jiang, P. B. Sujit, Srikanth Saripalli

TL;DR

This paper demonstrates that calibrating LiDAR intensity data for range and angle of incidence significantly improves semantic segmentation in challenging off-road environments. By treating intensity as a proxy for surface reflectivity, the authors outperform complex deep learning benchmarks in identifying difficult classes like "puddle" and "grass."

What Problem Does This Paper Address?

Autonomous navigation in off-road settings is notoriously difficult compared to urban environments. Off-road scenes lack structured geometry like curbs or lane markings and are filled with diverse objects that have "undefined boundaries."

Most state-of-the-art LiDAR segmentation models rely heavily on geometric features (the shape and arrangement of points). However, in the wild, a flat patch of grass and a flat puddle look identical to a sensor that only considers geometry. While LiDAR sensors also record "intensity"—the strength of the laser pulse reflection—this value is highly unstable. It fluctuates based on how far away an object is and the angle at which the laser hits the surface, making raw intensity unreliable for direct classification.

Methods

The authors propose a pipeline to transform raw, noisy intensity into a stable "reflectivity" value that can serve as a material signature.

  1. Intensity Calibration: Using the physical principle that received power ($P_r$) is proportional to $\frac{\rho \cos(\alpha)}{R^2}$ (where $\rho$ is reflectance, $\alpha$ is the angle of incidence, and $R$ is range), the authors developed a correction model.
  2. Handling the "Near-Range Effect": They observed that at very close distances (under 6 meters), the relationship between distance and intensity is non-linear and exponential. Their analysis focuses on the 6m–60m range where the physics is more predictable.
  3. Predicting the Angle of Incidence ($\alpha$): Calculating the angle of incidence from surface normals is often inaccurate using standard geometric methods. To solve this, the authors trained a Fully Connected Network (FCN) to predict a more accurate $\alpha$ using surface normal vectors as input.
  4. Cross-Platform Translation: Different LiDAR manufacturers process intensity differently. The authors introduced a polynomial fitting function, $Q(r)$, to convert Velodyne’s pre-processed 8-bit intensity data back into a raw format compatible with Ouster sensors, ensuring their method works across different hardware.
Raw vs. Calibrated Intensity for Grass

Datasets

The study utilizes the RELLIS-3D dataset, a multimodal off-road dataset containing annotated LiDAR point clouds.

  • Training/Calibration: Sequence 0000 from an Ouster OS1 (64-channel) sensor.
  • Testing: Sequences 0001 and 0002.
  • Cross-Platform Validation: Data from a 32-channel Velodyne Ultra Puck was used to test the intensity conversion methodology.

Key Insights and Ideas

  • Reflectivity as a Material Fingerprint: Once intensity is calibrated, different materials occupy distinct, predictable ranges. As shown in the violin plots below, classes like "puddle" and "grass" become much easier to separate.
  • Overcoming Geometric Ambiguity: For features like puddles, which have no height or distinct shape, intensity is the primary distinguishing factor.
  • The Power of Simplicity: The authors show that a relatively simple "neighborhood prediction policy" based on calibrated intensity can beat massive deep learning models in specific categories.
Calibrated Intensity ranges for different classes

Key Results

The proposed method was compared against two heavy-duty benchmarks: SalsaNext and KPConv.

  • Superior Puddle Detection: The authors achieved an Intersection over Union (IoU) of 47.83 for the "puddle" class, more than doubling the performance of SalsaNext (23.20) and vastly outperforming KPConv (0.0).
  • Improved Grass Segmentation: Their method also edged out benchmarks for "grass" with an IoU of 66.44.
  • Visual Accuracy: Qualitative results show that the intensity-based approach produces much sharper boundaries for ground-level features compared to traditional methods.
Segmentation Results on RELLIS-3D

Limitations and Future Work

  • Class Overlap: The "bush" class performed poorly (13.65 IoU) because its intensity range overlaps significantly with "trees." This is likely because both share similar biological textures (leaves and bark).
  • Environmental Sensitivity: Surface reflectivity changes with weather. Wet grass has a different signature than dry grass. The authors note that future work should evaluate the model across different climatic conditions.
  • Hybrid Integration: This paper serves as a "preliminary analysis." The ultimate goal is to integrate these calibrated intensity values as a standard input feature into more complex, learning-based semantic segmentation frameworks to get the "best of both worlds" (geometry + reflectivity).

Summary updated

Abstract

LiDAR is used in autonomous driving to provide 3D spatial information and enable accurate perception in off-road environments, aiding in obstacle detection, mapping, and path planning. Learning-based LiDAR semantic segmentation utilizes machine learning techniques to automatically classify objects and regions in LiDAR point clouds. Learning-based models struggle in off-road environments due to the presence of diverse objects with varying colors, textures, and undefined boundaries, which can lead to difficulties in accurately classifying and segmenting objects using traditional geometric-based features. In this paper, we address this problem by harnessing the LiDAR intensity parameter to enhance object segmentation in off-road environments. Our approach was evaluated in the RELLIS-3D data set and yielded promising results as a preliminary analysis with improved mIoU for classes "puddle" and "grass" compared to more complex deep learning-based benchmarks. The methodology was evaluated for compatibility across both Velodyne and Ouster LiDAR systems, assuring its cross-platform applicability. This analysis advocates for the incorporation of calibrated intensity as a supplementary input, aiming to enhance the prediction accuracy of learning based semantic segmentation frameworks. https://github.com/MOONLABIISERB/lidar-intensity-predictor/tree/main
BibTeX
@inproceedings{viswanath2023off,
  author = {Kasi Viswanath and Peng Jiang and P. B. Sujit and Srikanth Saripalli},
  title = {{Off-Road LiDAR Intensity Based Semantic Segmentation}},
  booktitle = {Iser},
  year = {2023},
  doi = {10.1007/978-3-031-63596-0\_54},
}

Loading related content...

Figure