Synthetic Generation of Radar Sensor Data for Virtual Validation of Autonomous Driving

The first accidents in otherwise promising deployments of autonomous driving fleets has underscored the importance of safety certification. Safety certification is expensive, especially when conducted via real-world driving. As such, high expectations are placed on virtual testing as an economic alternative. Autonomous driving functionality often relies heavily on radar sensors, but adequately modeling these radar sensors presents a particular challenge. While automotive simulation techniques have improved, there have yet to be systematic evaluations to prove that radar simulation models describe typical radar anomalies adequately such that downstream data processing algorithms behave correctly when operated with synthetic data. This dissertation builds a series of contributions to address this need. First, sensing artifacts are extracted from targeted measurements and evaluated for modeling relevance. A new method based on ray tracing is presented that generates spectral radar data from a 3D virtual environment, while addressing measurement ranges, limited resolution, and unambiguous intervals. This method, called "Fourier tracing", is of particular note in that it can model multipath related sensing artifacts. This dissertation presents a standard set of experiments for evaluating radar models and applies these to compare Fourier tracing to other modeling approaches using real radar data. The evaluation focuses on multipath propagation with respect to wave superposition, the visibility of occluded objects, on mirror targets, and on object tracking with synthetic data. The analysis makes it possible to falsify aspects of the examined radar sensor models, question their underlying assumptions, and to identify where these methods adequately simulate radar behavior