An Extreme Value Approach to Road Safety Analysis

Sammanfattning: In this thesis we study the feasibility of applying extreme value theory to data regarding road safety. In particular, we propose a model for assessing the risk of collision and near collision using extreme value theory. The thesis is relevant for road safety analysis in order to both understand whether extreme value theory is useful for modelling the collected data and to check if there is a need for collecting more data in future. Collecting this kind of information is very time consuming and expensive so efficient use of data is essential in this type of applications. The data consists of Time to Collision (TTC) and Post-Encroachment Time (PET) for right turning vehicles against bicycles in a four-way intersection in Barcelona. The dataset is the result of a 24-hour film sequence. The modeling is done with Generalized Extreme Value distribution and Generalized Pareto distribution with block maxima and peaks over threshold method. In addition, a homogeneous Poisson process model is suggested to make predictions on the number of collisions/near collisions in a longer time frame than the observed period.