Research Papers (2009 – 2013)

Abstract

Road user interactions develop continually over time and space, sometimes into more severe events such as conflicts and even collisions. Interactions can be considered as a fundamental event and all have the potential to develop into a collision. Evaluating how “close” road users were to an actual collision is performed using operational measures called surrogate safety measures such as Time to Collision (TTC). This paper relies on video analysis to extract road user trajectories from which several positional and safety indicators are then automatically derived. This work goes beyond current practice and research that relies on only one indicator value at a given time to qualify the whole interaction, e.g. the minimum TTC: the whole time series or profiles of the indicators are analyzed to find similarities between interactions with and without a collision. Data mining techniques are used to group interactions into homogeneous groups with similar indicator profiles. Spectral clustering is investigated as it does not rely on the distribution of data points like traditional clustering methods (e.g. k-means) but it depends on the eigen-decomposition of a similarity matrix. One key component in clustering is the similarity measure. In order to deal with profiles of various lengths without pre-processing that may distort the data, the Longest Common Subsequence similarity measure is used. The approach is demonstrated on a large real world dataset of traffic conflicts and collisions.

Mohamed Gomaa Mohamed, Nicolas Saunier