Road accident risk prediction using generalized regression neural network optimized with self-organizing map.

A road accident risk map that is designed to locate high-risk areas is an efficient way to reduce road traffic injuries and fatalities. To produce an accurate road accident risk map, it is required to compute the probability of the accident's occurrence by considering the various variables that contribute to road accidents. To this end, this research proposed a generalized regression neural network tuned with self-organizing map to estimate the risk of road accidents. This hybrid predictive model estimates the road accident risk by considering 22 different predictor variables (features), including geographic characteristics, temporal conditions, weather conditions, road-related characteristics, vehicle-related characteristics, and driver characteristics calculated based on the authoritative data sources and volunteered geographic information (VGI). The required VGI was collected in this study by developing a third-party application that was run inside telegram messenger. To evaluate the performance and usability of the proposed model for estimation of the accident risk along the road, the developed model was applied to the Tabriz-Marand dual carriageway, Iran. In this sense, 30 different scenarios were designed, and for each scenario, the risk of the accident was predicted at 3008 points along the Tabriz-Marand dual carriageway. A quality assessment of the proposed approach for different scenarios demonstrated that the predicted accident risk had very high accuracy (average accuracy about 90.74%). According to the results of this research, distance from traffic control cameras, day of the week, driver's age, weather, elevation, and vehicle type were the most effective factors in high-risk areas of the study area. [ABSTRACT FROM AUTHOR]