A Methodology for Safety Assessment of Highly Automated Vehicles: From Scenario Classification to Corner Case Identification / submitted by DI. Jinwei Zhou

Autonomous driving represents a technological leap forward that will likely solve key aspects of the transport problem and so have beneficial effects for some ciritical social and ecological issues as well. Indeed, thanks to increasing levels of automation, the Highly Automated Vehicle (HAV) is expected to be able to handle more and more complex traffic situations so that it can take over driving tasks and free human beings from driving for long periods of time. It is also expected to reduce accidents and improve traffic fluidity, reducing the overall carbon footprint, among other advantages. However, the higher the automation level, the more complex the HAV. As addressed by many, testing and validation of the HAV, so as to guarantee safety, is one of the most challenging tasks that still prevent the HAV from commercial release. Complexity excludes the use of well established methods for testing of classical vehicles, by on-road testing. This has led to a wide consensus that simulation “virtual testing” must be included as early as in the design phase. However, even simulation cannot test all possible situations. As HAVs will include many functions which can be updated frequently and require re-evaluation in short time, fast re-evaluations will be needed to prevent new dangers arising from the updates. This leads to the dilemma of HAV testing: on one side, we need to test enough situations to be confident about the behavior of HAV in the general, unknown case, but on the other side these tests need to be as limited as possible. Against this background, this research work is focused on developing such a methodology. The key idea is to replace on-road testing by accident statistics – so to say “involuntary” fleet testing – and to use model based methods as well as Design of Experiments to determine a limited set of cases to be tested. As an example to explain our approach, we concentrate on highways and starts by establishing a catalogue of countable scenarios, which cover over 90% of the (near-) crashes reported in U.S. highway databases between 2010 and 2013. Utilizing the catalogue, an approach has been developed leading to a realistic but parsimonious model based parametrization of the scenarios using experimental measurements. This allows covering the majority of the measured real traffic situation with a rather simple parameter set. The parametrization can then be used to determine a boundary that separates safe conditions from unsafe ones by a suitable Design of Experiment strategy.