Advancing LiDAR Intensity Simulation Through Learning With Novel Physics-Based Modalities

LiDAR sensors are integral to autonomous systems, providing a three-dimensional understanding of the surroundings. The intensity of LiDAR returns offers valuable information about the reflected laser signals, which facilitates crucial tasks such as object detection, classification, and segmentation. However, current physics-based LiDAR simulations fail to produce realistic intensity data. This research addresses this issue by using learning-based methods for realistic LiDAR intensity simulation. We propose a hybrid approach that incorporates novel physics-based modalities, specifically incidence angle and material reflectance, into the learning model to generate more realistic intensity data. We test this methodology across two architectures: (i) U-NET (Convolutional Neural Network) and (ii) Pix2Pix (Generative Adversarial Network), using the SemanticKITTI and VoxelScape datasets. The experiments compare the simulated intensity data generated by our method with state-of-the-art approaches through both qualitative and quantitative evaluations, and assessments of its effectiveness in improving the downstream tasks. The results demonstrate consistent improvements after the inclusion of the physics-based modalities.