Scenario-based trajectory generation and density estimation towards risk analysis of autonomous vehicles

Paper i proceeding, 2023

A large amount of testing is needed to determine when autonomous vehicles are sufficiently safe. To achieve this goal, test cases should be representative of real-world driving but also designed to provide sufficient coverage of both frequent and rare events. This is a crucial step in finding potential high consequence events, failure borders of the Autonomous Driving (AD) function and accurate estimation of the corresponding residual risks.

In this paper, we propose a new method to adapt generative models to generate vehicle trajectories that are representative of the ones collected from the real world. The method uses Non-Uniform Rational B-Splines (NURBS) combined with normalizing flows to build a statistical scenario model. The method allows us to estimate a joint probability density that can be used to evaluate the likelihood of different trajectory occurrences.

We demonstrate the method for statistical modeling on the (smooth and abrupt) cut-in traffic scenario and we give an example of how the estimated joint probability distribution can be used to assess the risk (trajectory occurrence probability and criticality) for different test cases. The results can be used for accelerated testing purposes, where the aim is to sample the rare tests more frequently, but can also be used to calculate the failure probability of AD functions.