Driver behavior models for evaluating automotive active safety: From neural dynamics to vehicle dynamics

Doctoral thesis, 2015

The main topic of this thesis is how to realistically model driver behavior in computer simulations of safety critical traffic events, an increasingly important tool for evaluating automotive active safety systems. By means of a comprehensive literature review, it was found that current driver models are generally poorly validated on relevant near-crash behavior data. Furthermore, competing models have often not been compared to one another in actual simulation. An applied example, concerning heavy truck electronic stability control (ESC) on low-friction road surfaces (anti-skidding support), is used to illustrate the benefits of simulation-based system evaluation with a driver model, verified to reproduce human behavior. First, a data collection experiment was carried out in a moving-base driving simulator. Then, as a complement to conventional statistical analysis, a number of driver models were fitted to the observed steering behavior, and compared to one another. The best-fitting model was implemented in closed-loop simulation. This approach permitted the conclusion that heavy truck ESC provides a safety benefit in unexpected critical maneuvering, something which has not been previously demonstrated. Furthermore, ESC impact could be analyzed at the level of individual steering behaviors and scenarios, and this impact was found to range from negligible, when the simulated drivers managed well without the system, to large, when they did not. In severe skidding, ESC reduced maximum body slip in the simulations by 73 %, on average. Some specific ideas for improvements to the ESC system were identified as well. As a secondary applied example, an advanced emergency brake system (AEBS) is considered, and a partially novel approach is sketched for its evaluation in what-if resimulation of actual recorded crashes. A number of new insights and hypotheses regarding driver behavior in near-crash situations are presented: When stabilizing a skidding vehicle, drivers were found to employ a rather simple and seemingly suboptimal yaw rate nulling strategy. Collision avoidance steering was found to be best described as an open-loop steering pulse of constant duration, regardless of amplitude. Furthermore, by analysis of data from test tracks as well as real-life crashes and near-crashes, it was found that detection of a collision threat, and also the timing of driver braking or steering in response to it, may be affected by a combination of situation kinematics and processes of neural evidence accumulation. These ideas have been tied together into a modeling framework, describing driving control in general as constructed from intermittent, ballistic control adjustments. These, in turn, are based on overlearned sensorimotor heuristics, which allow near-optimal, vehicle-adapted performance in routine driving, but which may deteriorate into suboptimality in rarely experienced situations such as near-crashes.