High-fidelity LiDAR Simulation System Based on Real Pointcloud Features

Virtual test is gaining increasing attention in the field of automated vehicles, and sensor models is a critical component in simulation. Despite LiDAR being an essential perception sensor equipped extensively on automated vehicles, the existing LiDAR pointcloud simulation software suffers from low fidelity in intensity, position and point dropout noise. To address this issue, we propose a system to analyze real pointcloud features and to simulate realistic pointcloud based on these features. Parameters of material properties are trained based on the statistical information of the collected pointcloud, and used in a pointcloud simulation pipeline to simulate intensity, position bias and detection possibility. The intensity expectation of simulated pointcloud is determined by a hybrid model which combines a data-driven Bidirectional Reflectance Distribution Function with a Gaussian model. Both the position bias and the detection possibility are calculated by intensity-related functions. Fidelity comparison is made between the pointcloud simulated by our proposed system, the VTD original method and the Learning to Predict LiDAR Intensity network, abbr. LPI, which is a state-of-the-art method. The fidelity of pointcloud from our proposed system is up to 32% higher than LPI.