Methods for Analysis of Naturalistic Driving Data in Driver Behavior Research

In the last several years, the focus of traffic safety research—especially when performed in association with the automotive industry—has shifted from preventing injury during a crash to avoiding the crash altogether or mitigating its effects. Pre-crash safety measures include intelligent safety systems (e.g., different levels of automated driving), infrastructure design, behavior-based safety, and policy-making. Understanding driver behavior is crucial in the development and evaluation of such measures. Naturalistic driving data (NDD) can facilitate this understanding by providing information about crash causation and contribute to the evaluation of pre-crash safety measures and the effects of driver behavior on safety. However, NDD’s complexity calls for new and better methods to fully exploit its advantages. This thesis, together with the five included papers, addresses several gaps in current scientific knowledge by presenting novel methods for analyzing NDD that address multiple aspects of the development process for pre-crash safety measures. The chunking method (Paper I) helps to identify and overcome common biases in analysis of everyday-driving time-series data, while the expert-assessment-based crash-causation analysis method (Paper II, supported by Paper III) is a novel approach to studying crash causation through the analysis of NDD with video. Product and prototype development can be improved by utilizing counterfactual simulations, for which the choice of driver behavior model is shown to be crucial (Paper IV)—an awareness that was previously lacking. Being able to compare the effects of drivers’ specific behaviors (e.g., driver-vehicle interactions or in-vehicle secondary tasks) on safety could both speed up development of safety measures and improve vehicle designs and design guidelines. Methods to perform such comparisons through the combination of counterfactual glance behavior and pre-crash kinematics have been missing (they are provided in Paper V). This thesis further improves the evaluation of pre-crash safety measures by providing more robust analyses of everyday driving data (Paper I) and by demonstrating the importance of good mathematical models of driver behavior in virtual evaluation (Paper IV). In summary, these new methods fill important research gaps and have the potential to improve the design of pre-crash safety measures through the use of NDD. Using NDD can augment our understanding of driver behavior and crash causation, important aspects of improving traffic safety and fulfilling Sweden’s Vision Zero