Clustering Surrogate Safety Indicators to Understand Collision Processes
Nicolas Saunier, Mohamed Gomaa Mohamed · PolyPublie (École Polytechnique de Montréal) · 2014
As time series are collected through more and more pervasive devices carried by users and vehicles, new tools are necessary to understand and mine the large amounts of transportation data being thus generated. This work proposes a new similarity measure for time series that is applied to surrogate measures of safety and other indicators characterizing road user interactions. The new similarity measure based on the aligned longest common sub-sequence is paired with a custom clustering algorithm that does not require to set the number of expected clusters and remains interpretable through the use of prototype indicator profiles as cluster representatives. The method is applied to five indicators, including time to collision and probability of collision, for a large real world dataset of traffic videos of collisions and conflicts. The results confirm the general assumption of surrogate methods for safety analysis that some interactions without a collision have very similar processes to collisions. It also highlights the danger of using a significant proportion of candidate interactions without a collision that seem to share little similarities with collisions.