How to evaluate synthetic radar data? Lessons learned from finding driveable space in virtual environments

Generating synthetic sensor readings at scale by means of virtual sensors is expected to facilitate safety validation of autonomous driving functions. An absolute equality of real and synthetic data is not to be expected. Instead, it has to be proven that synthetic sensor data exhibits a comparable level of uncertainties as data from the real sensor, so that subsequent algorithms draw the same conclusions from the respective input data. This paper addresses this problem by comparing free space information inferred from real and synthetic radar data. It is shown that comparable free space can be calculated from the sensor simulation, although deviations between synthetic and real sensor data exist. The presented method to compare the calculated free space from synthetic and real data serves as evaluation of the simulation model.