Geo-referenced Data and Complex Networks for Measuring Road Accident Risk

The assessment of risk related to car crashes in road networks is a relevant topic for both social and political decisions and for insurance companies. To this end, we show how the spatial objects and the information concerning the structure of the roads, that can be collected e.g. from open data sources, along with the crash history can be used to map the risk related to each road. In particular, we follow a combined approach. On the one hand, a statistical model is developed in order to assess the risk on the basis of a set of features related to the characteristics of the streets. On the other hand, from the spatial object we build a weighted network, where vertices and arcs correspond to geographical elements as junctions and roads respectively and where the assessed risk of each segment is used as a weight. We study the topology structure of the graph obtained and we show how classical network indicators can provide meaningful insights about the risk of an area.

To achieve our aim, we need to adapt the current methodology about geospatial modelling to the constraints derived from the maps of the roads of a particular area and to exploit supervised/unsupervised statistical learning algorithms to estimate the local risk of the frequency of accidents. We do not consider here other features that can be detected by telematic data or by adding other data sources (e.g., driving behaviour, driving habits, KM coverage, daytime, weather conditions, etc.).

A statistical model is developed in order to assess the risk on the basis of a set of features related to the characteristics of the roads. We split the domain into disjoint subregions and we apply separately the model on each subregion to appreciate the specificity of each area. Additionally, to include spatial dependence, we take into account in the model the features of segments that are in the proximity of the road object of interest. Each of this feature is weighted using the inverse of a distance measured on the network. The distance is here based on the concept of weighted shortest part.

The aim of this approach is indeed to get the model that assures the best adaptation to the peculiar details of an area and, at the same time, is not affected too much by details and characteristics observed far from the spatial domain of interest.

The spatial object and the accident risk assessed by the model for each road are then converted in a directed and weighted graph. In particular, we focus on a “junction graph", where each segment is an arc and nodes are given by junctions (or by termination of closed streets). Each arc is then weighted according to the risk of the segment detected at previous step. Focusing on network topological indicators, we observe a significant correlation between the risk associated to a node and the node betweenness measured on the network. Therefore, the centrality of a node in the topological structure appears related to the risk measured by the model. Additionally, by means of the Louvain methodology, we detect communities in the area. The communities depend on both the arc density and on the weights. The split of the area into clusters can be used by insurance companies to measure the propensity to get an accidents in the neighbour of a point, and then to fine tune the cost of premiums to be paid to drive a car. A numerical application based on Milan area in Italy (city and province) is provided.