Simulation-Based Elicitation of Accuracy Requirements for the Environmental Perception of Autonomous Vehicles.

Novel methods for safety validation of autonomous vehicles are needed in order to enable a successful release of self-driving cars to the public. Decomposition of safety validation is one promising strategy for replacing blunt test mileage conducted by real world drives and can be applied in multiple dimensions: shifting to a scenario-based testing process, assuring safety of individual subsystems as well as combining different validation methods. In this paper, we facilitate such a decomposed safety validation strategy by simulation-based elicitation of accuracy requirements for the environmental perception for a given planning function in a defined urban scenario. Our contribution is threefold: a methodology based on exploring perceptual inaccuracy spaces and identifying safety envelopes, perceptual error models to construct such inaccuracy spaces, and an exemplary application that utilizes the proposed methodology in a simulation-based test process. In a case study, we elicit quantitative perception requirements for a prototypical planning function, which has been deployed for real test drives in the city of Hamburg, Germany. We consider requirements regarding tracking and the position of an oncoming vehicle in a concrete scenario. Finally, we conclude our methodology to be useful for a first elicitation of quantifiable and measurable requirements.