Computational driver behavior models for vehicle safety applications

Sammanfattning: The aim of this thesis is to investigate how human driving behaviors can be formally described in mathematical models intended for online personalization of advanced driver assistance systems (ADAS) or offline virtual safety evaluations. Both longitudinal (braking) and lateral (steering) behaviors in routine driving and emergencies are addressed. Special attention is paid to driver glance behavior in critical situations and the role of peripheral vision. First, a hybrid framework based on autoregressive models with exogenous input (ARX-models) is employed to predict and classify driver control in real time. Two models are suggested, one targeting steering behavior and the other longitudinal control behavior. Although the predictive performance is unsatisfactory, both models can distinguish between different driving styles. Moreover, a basic model for drivers' brake initiation and modulation in critical longitudinal situations (specifically for rear-end conflicts) is constructed. The model is based on a conceptual framework of noisy evidence accumulation and predictive processing. Several model extensions related to gaze behavior are also proposed and successfully fitted to real-world crashes and near-crashes. The influence of gaze direction is further explored in a driving simulator study, showing glance response times to be independent of the glance's visual eccentricity, while brake response times increase for larger gaze angles, as does the rate of missed target detections. Finally, the potential of a set of metrics to quantify subjectively perceived risk in lane departure situations to explain drivers' recovery steering maneuvers was investigated. The most influential factors were the relative yaw angle and splay angle error at steering initiation. Surprisingly, it was observed that drivers often initiated the recovery steering maneuver while looking off-road. To sum up, the proposed models in this thesis facilitate the development of personalized ADASs and contribute to trustworthy virtual evaluations of current, future, and conceptual safety systems. The insights and ideas contribute to an enhanced, human-centric system development, verification, and validation process. In the long term, this will likely lead to improved vehicle safety and a reduced number of severe injuries and fatalities in traffic.