Many existing geostatistical methods for defining concentrations, such as the kernel and the local spatial autocorrelation method, take into account the Euclidian distance between the observations. However, since traffic accidents are typically located along a road network, it is assumed that the use of network distances instead of Euclidean distances could improve the results of these clustering techniques. A new method is proposed here, taking into account the distances along the road network. This methodology is applied to and tested on the historic city of Brussels (Belgium). Most existing accident clustering methods start from the actual distribution of the accidents over the road network. However, to guarantee the independence between the results and the actual accident locations, a close distribution of points of measurement is randomly distributed over the road network. A cost matrix is calculated, containing the distances between each individual accident and the points of measurement. Next, a dangerousness index is calculated for each point of measurement, taking into account the weight of the accidents. Nearby accidents have a greater influence on the dangerousness index than accidents further away. Finally, a network Voronoi diagram is used to distribute the dangerousness index over the road network. The proposed methodology allows computing spatial concentrations of road accidents based on distances along the road network, including distances between accidents on different, but intersecting roads. By using randomly distributed points of measurement instead of accident locations, the independency between the accident locations and the resulting dangerousness segments is guaranteed, allowing for easy comparison of results in different analyses (for example in different time periods). For the covering abstract see ITRD E138952. This paper is available from http://www.ictct.org/workshops/06-Minsk/Aerts.pdf.
Abstract