Statistical modelling of critical cut-ins for the evaluation of autonomous vehicles and advanced driver assistance systems

Abstract

Understanding human behaviour in traffc is an integral part of developing active safety systems (ADAS) and autonomous vehicles (AV). Such systems require rigorous testing before they can be put in commercial use. This thesis aimed to study data collected as part of the second Strategic Highway Research Program (SHRP2) Naturalistic Driving Study (NDS) to improve the estimation made in the tails of a driver model for lane changing. This was to be done through annotating a data-set of 1191 critical lane-change events provided through SHRP2 NDS. Annotation was done using an annotation tool that was developed for this purpose as part of a previous project. The trajectories of the manoeuvres were then extracted and parameterised using ridge regression. It was found that 86 of the 1191 events were suitable for annotation. Due to the limited quality of the data and number of usable events, the thesis aim was redirected to model the uncertainty of the annotation method using 9 events annotated by 5 annotators. Two linear regression models were then developed to estimate the uncertainty of this annotation method. The results show that the models can predict the uncertainty based on the limited number of events that were available. These results have potential to be used to estimate the uncertainty of the parameterised trajectories in future work.