Method Development of Multi-Dimensional Accident Analysis Using Self Organizing Map

Implementation of appropriate safety measures, either from the viewpoint of a vehicle or the society or the infrastructure, it is an important issue to clearly understand the multi-dimension complicated real world accident scenarios. This study proposes a new method to easily capture and to extract the essence of such complicated multi-dimension mutual relationship by visualizing the results of SOM (Self Organizing Map). The FARS data from 2010 is used to generate a dataset comprised of 16,180 fatal passenger car drivers and 48 variables. The 16,180 fatal drivers were clustered using hierarchy cluster analysis method and mapped into a two-dimensional square with one dot representing one fatal driver using SOM. From the SOM assessment of the 16,180 fatal drivers, five clusters were created, and they are characterized as follows: Cluster 1 (Interstate highway accidents), Cluster 2 (Drunk speeding), Cluster 3 (Non speeding lane departure), Cluster 4 (Vehicle to vehicle) and Cluster 5 (Intersection). The number of fatalities in Clusters 1 and 2 could be possibly reduced by application of CA (Crash Avoidance) technologies and stricter enforcement of traffic laws. For Clusters 3, 4 and 5, reduction by CA technologies and stricter enforcement would be more difficult, because (i) the majority of the drivers would be respecting the laws and (ii) the road environments and vehicle to vehicle situations are more complicated for current CA technologies. The remaining crash scenarios related to Clusters 3, 4 and 5 are head on collision / impact to fixed object on minor roads and side impact in urban areas.