An Approach to Predict Road Accident Frequencies: Application of Fuzzy Neural Network

Road accident prediction plays an important role in accessing and improving the road safety. Besides the conventional generalized linear regression, the prediction approaches based on fuzzy logic and neural networks have increasingly been proven to have a significant accident-predicting capability in recent years. However, fuzzy logic and neural network have their respective limitations. For example, it is difficult to construct a complete rule set for fuzzy logic and there is no general rule in determining the network structure for neural network. To overcome these limitations, the fuzzy neural network (FNN) is put forward. This approach has been applied for prediction in many areas, but no application exists in road accident prediction according to the authors’ knowledge. Thus, this paper establishes a fuzzy neural network model (FNNM) for predicting accident frequencies. It is established based on a data set of 133 segments from urban arterials in Harbin city of China, which takes annual average daily traffic (AADT), lane width (LW), speed limit (SL) and traffic load (TL, calculated by volume/capacity) as input variables and accidents per kilometer per year (AF) as output variable. Comparisons among FNNM, fuzzy logic model (FLM) and BP neural network model (NNM) show the superiority of the FNNM in accuracy and flexibility. Finally, a sensitivity analysis is employed to identify the significant factors. The results show that AADT is the most significant factor in this model, followed by SL, TL and LW in order of their relative importance going from the most to the least significant.