Safety in Mixed Traffic Environment

Safety is an essential target of the automation of driving functions. Safety is usually understood mainly as primary safety, avoiding being involved in accidents, but secondary safety is important as well, i.e., not inducing risks to other vehicles. While primary safety is always considered in Advanced Driver Assistance Systems (ADAS) development, secondary safety is usually not. The key question of this thesis: should it be considered? and if so, how? To answer these questions, we first start by noting that automation is taking place mainly in form of single ADAS. While it is widely expected that this will eventually lead to a fully automated vehicle, in the meantime partly or even fully automated vehicles will have to share the driving space with human-driven vehicles. To test our ideas, we concentrate on an example, a highway merging assistant, which is potentially one of the most dangerous situations on highways. We then compare two versions of this ADAS, one which considers only the safety of the controlled vehicle and another one that tries to minimize the disturbance and so the risk to the next participants. The results show that it indeed makes sense to include secondary safety in the design of the ADAS algorithm, but even more for another aspect, the fluidity of traffic. Including secondary safety means asking our ADAS to avoid behavior that might be unexpected and/or disturbing for human-driven vehicles and lead to a risky reaction. Defining such acceptable behavior is not a trivial task, as decisions by a human driver will depend on many factors, some of them time-varying but quantifiable ones, like weather and traffic conditions, but also very much on less crisp ones like the local habits. Against this background, this thesis proposes a double layer model of human driving behavior, in which a first decision level can be tuned to the actual situation while a second actuation level is independent of it. Data sets from different countries and driving conditions confirm the plausibility of this hypothesis. Then acceptable behavior can be defined in several ways but essentially as the inversion of the model.