Generation and analysis of driving scenario trajectories for safety verification of Autonomous Drive

Abstract

Autonomous vehicles are often considered to be the future of transportation. However, many believe fully-autonomous vehicles are still far away from being realized. One of the main reasons is the lack of confidence from a safety level standpoint. While car manufacturers such as Volvo Cars have made great efforts to prove that developed algorithms are safe enough, it is just not feasible to test each new feature in the fields. Luckily, virtual environments address this issue by reproducing traffic situations - scenarios. But even with virtual environments, a lot of data still has to be collected to cover the whole spectrum of possible scenarios. This can be addressed with simulated scenarios. Another approach that is investigated in this thesis is generating scenarios based on real ones. We considered two approaches: Recurrent Conditional Generative Adversarial Nets and a Recurrent Autoencoder with Generative Adversarial Nets. The latter also shows useful properties for scenario analysis. The second problem that we try to tackle in this thesis is the evaluation of our developed models.