Physical and phenomenological modeling of automotive LiDAR sensors for real-time applications

Automotive LiDAR sensors play a crucial role in environment perception and mapping for automated driving systems. Vehicle-in-the-loop (ViL) testbenches offer a safer, more controlled, and less costly alternative to on-road testing, but they require LiDAR sensor models that are both physically accurate and real-time capable. Meeting these requirements is challenging: laser-surface interactions must be reproduced for a wide variety of materials, motion-induced distortions must be rendered on the fly, and weather phenomena such as rain and snow must be represented realistically, all while maintaining real-time performance. Current LiDAR simulators, such as those in Carla and AirSim, lack validated reflectance models, struggle to accurately simulate the motion distortion effect in real time, and do not account for precipitation effects. This dissertation addresses these shortcomings through three main contributions: First, a hybrid approach for modeling LiDAR point cloud reflectance is proposed. The method uses bidirectional reflectance distribution function (BRDF)-based models to estimate the reflectance of primary scene materials, including road surfaces, traffic infrastructure, and vehicle coatings. The BRDF parameters are identified from controlled LiDAR measurements conducted in our laboratory. Quantitative evaluation confirms that the BRDF-based reflectance models accurately reproduce measured point cloud reflectances for the target materials. Secondary background elements, where analytical modeling is intractable, are represented using statistical reflectance priors derived from large-scale real-world datasets. Second, a novel approach named Trace One More Time (TOMT) was developed to simulate the motion distortion effect in real time for dynamic driving scenarios involving both static and moving objects. This method optimizes computational processes by correcting ray origins instead of repeatedly updating the acceleration structure. As a result, it significantly reduces computational complexity while preserving accuracy. Experiments confirm close alignment with reference data and validate TOMT's real-time suitability for ViL applications. Third, two specialized methods were devised to incorporate precipitation effects into simulated LiDAR point clouds. First, measured backscatter from precipitation is injected directly into the synthetic point cloud, with corresponding reflectance values estimated using a Mie scattering model. Second, multipath propagation caused by specular reflections on wet pavement is modeled analytically. Combined, these techniques enable the proposed LiDAR sensor model to generate point clouds that more closely resemble field LiDAR data than those produced by current benchmark models. A series of real-time capability tests confirmed that the LiDAR sensor model meets real-time requirements even in its worst-case configuration, demonstrating its suitability for applications such as ViL testing. Collectively, these advances provide a high-fidelity model for validating LiDAR-based automated driving systems and support the development of safer, more reliable autonomous vehicles.