Physically Accurate Lidar Simulation for Automotive Digital Twins

Recent advancements in real-time ray tracing show promise for creating digital twins in automotive applications. AI perception models can be validated and fine-tuned using digital twin sensor data, improving autonomous driving in edge cases. However, there is still a gap in the realism of these simulations. This paper addresses specifically the simulation of LiDAR by combining physics-based ray tracing from a graphics buffer with a database of real-world material properties captured in the infrared spectrum under various weather conditions. Compared to other approaches, our method synthesizes point clouds that match much closer those from commercial LiDAR scanners under similar conditions than pure geometry-based simulations (e.g. from CARLA). We demonstrate this through a performance study on pedestrian detection task, where a state-of-the-art model pre-trained on real data performs equally well on our synthesized data. Our digital twin LiDAR sensor works independently of existing simulation enviornments and can be used in single or multi-sensor simulations without additional tuning.