Publication

Automated vehicles (AVs) are expected to bring safer and more convenient transport in the future. Consequently, before introducing AVs at scale to the general public, the required levels of safety should be shown with evidence. However, statistical evidence generated by brute force testing using safety drivers in real traffic does not scale well. Therefore, more efficient methods are needed to evaluate if an AV exhibits acceptable levels of risk.

This thesis studies the use of two methods to evaluate the AV's safety performance efficiently. Both methods are based on assessing near-collision using threat metrics to estimate the frequency of actual collisions. The first method, called subset simulation, is here used to search the scenario parameter space in a simulation environment to estimate the probability of collision for an AV under development. More specifically, this thesis explores how the choice of threat metric, used to guide the search, affects the precision of the failure rate estimation. The result shows significant differences between the metrics and that some provide precise and accurate estimates.

The second method is based on Extreme Value Theory (EVT), which is used to model the behavior of rare events. In this thesis, near-collision scenarios are identified using threat metrics and then extrapolated to estimate the frequency of actual collisions. The collision frequency estimates from different types of threat metrics are assessed when used with EVT for AV safety validation. Results show that a metric relating to the point where a collision is unavoidable works best and provides credible estimates.

In addition, this thesis proposes how EVT and threat metrics can be used as a proactive safety monitor for AVs deployed in real traffic. The concept is evaluated in a fictive development case and compared to a reactive approach of counting the actual events. It is found that the risk exposure of releasing a non-safe function can be significantly reduced by applying the proposed EVT monitor.