Physically-Based Simulation of Automotive LiDAR

TL;DR

The paper addresses the need for realistic synthetic automotive LiDAR data, including blooming artifacts, by proposing a physically-based rendering framework that links emitter, scene, and receiver via a single-bounce optical model. It introduces two algorithmic approaches—beam iteration and range stacking—to compute LiDAR echoes from rasterized or ray-traced scenes, with parameters extracted from optical laboratory measurements. The authors demonstrate calibration on two automotive LiDAR systems and provide practical rendering considerations, offering a path to high-fidelity synthetic data for training and HiL testing while acknowledging computational limitations. This work advances domain-accurate synthetic LiDAR data generation and highlights concrete steps for parameterization and validation against real sensors.

Abstract

We present an analytic model for simulating automotive time-of-flight (ToF) LiDAR that includes blooming, echo pulse width, and ambient light, along with steps to determine model parameters systematically through optical laboratory measurements. The model uses physically based rendering (PBR) in the near-infrared domain. It assumes single-bounce reflections and retroreflections over rasterized rendered images from shading or ray tracing, including light emitted from the sensor as well as stray light from other, non-correlated sources such as sunlight. Beams from the sensor and sensitivity of the receiving diodes are modeled with flexible beam steering patterns and with non-vanishing diameter. Different (all non-real time) computational approaches can be chosen based on system properties, computing capabilities, and desired output properties. Model parameters include system-specific properties, namely the physical spread of the LiDAR beam, combined with the sensitivity of the receiving diode; the intensity of the emitted light; the conversion between the intensity of reflected light and the echo pulse width; and scenario parameters such as environment lighting, positioning, and surface properties of the target(s) in the relevant infrared domain. System-specific properties of the model are determined from laboratory measurements of the photometric luminance on different target surfaces aligned with a goniometer at 0.01° resolution, which marks the best available resolution for measuring the beam pattern. The approach is calibrated for and tested on two automotive LiDAR systems, the Valeo Scala Gen. 2 and the Blickfeld Cube 1. Both systems differ notably in their properties and available interfaces, but the relevant model parameters could be extracted successfully.

Figures (7)

• Figure 1: Included and excluded aspects of the single-bounce model: LiDAR rays originate at the emitter/collector E with intensity $\eta_b(\Phi)$, continue to a single target spot T, and return at the same angle with intensity $\rho_b(\Phi)$. Stray light travels from a source (e.g., the sun A) to T, then to E. Virtual sources C, (e.g., atmospheric scattering B), can also be considered. However, indirect paths from/to E, such as multiple bounces over the road I, are not modeled.
• Figure 2: Principle of range-separated blooming for first dominant echo. Lines Z in the bird's eye view indicate iso ranges. Signals along lines of sight (X, Y) do not mix additively across different ranges. Instead, the strongest echo is propagated, forming regions (A$'$--D$'$) in which retroreflections of one of the four front vehicles, and hence their ranges A--D, dominate. X and Y are located at spatial angles at which the echoes of the respective car and truck have matching intensity. Optical effects are exaggerated for visibility, and only the range dominance among A--D is compared.
• Figure 3: Comparison of increased resolution vs. the correction (through 10-bit normals) for reducing angular errors. It can be seen that even at high resolutions of $8192$ pixels wide, step artifacts remain distinct at far ranges. In contrast, when correcting via normals, residual steps exclusively arise from lack of precision in normals, if applicable. Increased resolution in this case also serves to eliminate extrapolation errors from erroneous normals.
• Figure 4: Effect of clipping planes and thus z buffer resolution with normals correction. A near plane of $10^{-4}$ and a far plane of $10^6$ provide fair results. Pulling the far plane to $10^2$ provides little improvement. Pushing the near plane up to $10^{-5}$, in turn, yields degenerate, discrete depth levels.
• Figure 5: Results of the optical characterization of the Blickfeld Cube 1. Each group of three spots corresponds to one beam direction, as the sensor employs three laser diodes to construct an approximately square pulse shape. Highly-resolved beam intensity in (b) is measured with retro-reflective screen. The inner section of the pulse shape marked with a white rectangle is scaled down by a factor of 1 000 to visualize the entire pulse within the same range. While offering the highest grade of retro reflectance, the prism optics introduce a higher dependency on the incidence direction and a speckled appearance. Dark stripes result from the horizontal application of the tape. A more uniform retro-reflective surface based on glass beads could be more suitable. Relative sensitivity in (c) over input angles is composed of separate horizontal and vertical measurements via $H * V$.