A Multidimensional Clustering Algorithm for Studying Fatal Road Crashes

Road fatalities are rare outcomes of events that occur in a small time-space region. Although the exact chain of events for each fatality is unique, there are inherent similarities between road fatalities. The science of road safety is dedicated to identifying such similarities, mainly using statistical analysis tools. Researchers typically analyze patterns that emerge over space, such as hot-spot studies, or patterns that emerge over time, such as before-after studies. Traffic research enumerates 84 parameters that characterize a road fatality. A vast number of papers have tried to find the correlation between one or two parameters. In those studies quite often the contribution of other factors is omitted. In this research we utilize a clustering graph theoretic method, known as graph-cuts, for segmenting a very large crash dataset (i.e., all fatal car crashes in the last 2, 5, or 10 years), while incorporating all available crash information into the process. The analysis of the clusters allows one to find subtle trends and significant causes for traffic fatalities. With this method, for example, we have found high correlation between hit-and-run and pedestrians fatalities, which was overlooked by previous studies. An additional output of the research is a full description of the typical fatality, thus all factors that characterized the representative crash in a cluster.