Detecting crash hotspots using grid and density-based spatial clustering

Abstract

Data mining techniques, specifically spatial clustering methods, are used to analyse crash data and find their spatial patterns. In the present study, a grid and density-based clustering algorithm called GriDBSCAN was utilised for injury crash data. Other clustering methods such as nearest neighbour hierarchical and kernel density estimation were also applied to validate the results of the GriDBSCAN algorithm. Crash points recorded for Gebze and Izmit (in Turkey) were clustered through these methods. The findings revealed that GriDBSCAN had the highest value for hit rate. In addition, the GriDBSCAN algorithm placed data points into a grid mesh to decrease the runtime and could estimate the clusters with a higher accuracy due to the recognition of the noise points. Furthermore, the proposed approach allowed the detection of unique crash factors for both cities. The factors contributing to injury crashes in both cities included collision and junction types, along with speed limit.