On Safety Validation of Automated Driving Systems using Extreme Value Theory
Autonomous vehicles are expected to bring safer and more convenient transports in the future. When the system in the vehicle takes care of the driving, the driver is free to spend time on other things. As the driver is no longer part of the loop and cannot be used as a fallback, the requirements that are put on safety and dependability of the system will be very high. To test the system in real traffic and measure the failure rate that leads to an accident will therefore not be feasible. However, due to the complexity of the system, it is still desirable to be able to test the safety on a complete system level.
With the emergence of automated driving systems, the vehicles will be equipped with an array of sensors that gives a representation of the environment. This opens up the possibility to use more information to estimate how safe the system behaves in real traffic. Using an area of statistics called Extreme Value Theory, the frequency of near-collision can be extrapolated into a frequency of actual collisions.
These near-collisions are measured using threat assessment methods that have been developed for active safety applications. In this thesis, two types of measures are evaluated to determine how well they can be used for extrapolation. From the results, it is clear that the measure relating to a point where a collision is unavoidable works better than the one relating to the actual collision.
Furthermore, several methods for automatically fitting the extreme value model to the data are evaluated. The result shows that all tested methods work well where some methods put emphasis on the more extreme data, which can result in a difference of the inferences drawn. This suggests that the whole process has the possibility to be automated, which is necessary when performed repeatedly on multiple large data sets.