On the analysis of naturalistic driving data
In the last several years, the focus of traffic safety research has shifted from injury prevention during a crash to measures taken before a crash, in order to mitigate its effects or avoid it completely. Measures include advanced driver assistance systems, safety aspects of autonomous driving and infrastructure design, behavior-based safety (driver training), and policy-making. All of these pre-crash measures require an understanding of driver behavior.
As a result of this need, naturalistic driving data (NDD) has emerged as a crucial data source with high ecological validity. NDD enable not only the real-world assessment of driver behavior, but also that of road infrastructure and pre-crash safety measures. However, NDD’s great potential is hindered by its complexity. Consequently, new methods to analyze NDD are greatly needed.
This thesis presents a novel framework for traffic safety research using NDD and discusses the framework’s benefits and drawbacks. Furthermore it presents novel methods for analyzing NDD. The first paper presents a robust method to reduce bias in the analysis of time-series NDD. The second paper ports the DREAM method, used in traditional on-scene crash investigations, to vehicle-to-pedestrian incidents in NDD with video data. The third paper analyzes NDD with a novel method based on expert judgment. This method, inspired by DREAM, is currently applied to commercially collected and event-based, real-world crashes with driver and forward video. Finally, the fourth paper presents a new, pragmatic method to extracting range, range rate and optical parameters (e.g. looming) from the forward video in commercially collected lead-vehicle NDD.
In summary, the methods developed and presented in this thesis use quantitative and qualitative analyses of time-series and video data from naturalistic driving to augment our understanding of driver behavior. Pre-crash safety measures will be further advanced not only by these insights, but also by future applications of the methods developed in this thesis.