Using Data Mining Techniques to Understand Collision Processes

In order to improve road safety, it is necessary to better understand collision processes, i.e. the chains of events that lead to collisions. Among the most important benefits, more efficient countermeasures can be found to target causes and factors known to lead to collisions. This would also help develop more reliable surrogate safety measures based on traffic events without a collision that have stronger links to collisions. This paper reports on the first phase of a project relying on microscopic data extracted from video sensors and on data mining techniques to identify patterns in a dataset of traffic events with and without a collision. This approach is demonstrated on a dataset collected in Kentucky of 295 traffic events, constituted of 213 conflicts and 82 collisions. Using the k-medoid algorithm, its clustering yields three groups with distinct characteristics, especially related to speed and the type, or lack, of evasive action. The most important attributes that determine the traffic event outcome (collision or not) are identified through a logit model.