Multivariate spatial clustering of traffic accidents for local profiling of risk factors

Abstract

According to previous studies, traffic accident data have a spatial dependence which should be taken into account when analyzed. For this purpose, a proper spatial segmentation of accidents should be carried out so that subsequent spatial analysis can provide significant results. In this work, we propose a method for spatial clustering of multiple variables in order to make a new spatial characterization of the different road stretches and then to assign them into a small set of typical accidents according to their risk profile. First, every road is segmented according to an estimation of the corresponding spatial accident density. Then, each segment is characterized with a numerical vector representing accident attributes. The spatial clustering is performed in the third stage by applying a k-means clustering algorithm. Traffic accident data from Comunidad Valenciana, in Spain, have been used for testing our method. Results show that our approach is a flexible and intuitive way for spatially characterizing the roads of the region under study, and even for finding relationships between values of the analyzed risk factors.