Much effort is put right now into how to make autonomous vehicles as capable as possible in order to be able to replace humans as drivers. Less focus is put into how to ensure that this transition happens in a safe way that we can put trust in. The verification of the extreme dependability requirements connected to safety is expected to be one of the largest challenges to overcome in the commercialization of autonomous vehicles. Using traditional statistical methods to validate complete vehicle safety would require the vehicle to cover extreme distances to show that collisions occur rare enough. However, recent research has shown the possibility of using near-collisions in order to estimate the frequency of actual collisions using Extreme Value Theory. To use this method, there is a need for a measure related to the closeness of a collision. This paper shows that the choice of this threat measure has a significant impact on the inferences drawn from the data. With the right measure, this method can be used to validate the safety of a vehicle. This, while keeping the validity high and the data required lower than the state of the art statistical methods.
Using Extreme Value Theory for Vehicle Level Safety Validation and Implications for Autonomous Vehicles
IEEE Transactions on Intelligent Vehicles ; 2 , 4 ; 288-297
2017-12-01
974284 byte
Article (Journal)
Electronic Resource
English
Interpretable Safety Validation for Autonomous Vehicles
IEEE | 2020
|Real-time safety analysis using autonomous vehicle data: a Bayesian hierarchical extreme value model
Taylor & Francis Verlag | 2023
|