Citation

Saunier N, Mourji N, Agard B. Transp. Res. Rec. 2011; 2237: 41-50.

Copyright

(Copyright © 2011, Transportation Research Board, National Research Council, National Academy of Sciences USA, Publisher SAGE Publishing)

DOI

10.3141/2237-05

PMID

unavailable

Abstract

Road collisions lead to great human and financial costs for society. Although some progress has been made, this worldwide issue needs more attention, as the costs increase. Proactive methods for road safety analysis that do not depend on collision occurrences are needed. Collection and analysis of microscopic data (road user trajectories) about all traffic events with and without a collision are the only ways to gain insight into collision factors and processes; that is, the chains of events that lead to collisions. The first phase of the project reported in this paper used microscopic data extracted from video sensors and data mining techniques to identify patterns in the traffic event database. Decision trees, the k-means algorithm, and the hierarchical agglomerative clustering method were used to analyze the relationship between interaction attributes and outcome (collision or not) and identify groups of interactions with similar attributes. This approach was demonstrated on a data set collected in Kentucky of 295 traffic events and contained 213 conflicts and 82 collisions. The decision tree confirmed the importance of evasive action in the interaction outcome. Three clusters were found from speed indicators extracted from road users' trajectories: the cluster containing the fewest collisions had the lowest speeds of the three. This result hints at the existence of conflicts that are dissimilar from most collisions and may therefore not be suitable for surrogate safety analysis.