Application of Adaptive Neuro-fuzzy Inference System for road accident prediction

Abstract

Since the last two decades, several modeling approaches have been developed in road safety literature to establish the relationship between traffic accidents and road characteristics. However, to the best of the authors’ knowledge, no extensive research work has been published on application of Adaptive Neuro-fuzzy Inference System (ANFIS) on road accident modelling. Therefore, the present paper aims to develop an ANFIS technique for modelling traffic accidents as a function of road and roadside characteristics. To achieve the objective, accident data and road characteristics were collected over a two-year period along the Qazvin-Loshan intercity roadway in Iran. The candidate set of explanatory variables included the Mean Horizontal Curvature (MHC), Shoulder Width (SW), Road Width (RW), Land Use (LU), Access Points (AP), Longitudinal Grade (LG), and Horizontal Curve Density (HCD). The results showed that RW, SW, LU, and AP significantly affected accident frequencies. Using statistical performance indices, the ANFIS model was compared with the Poisson, negative binomial, and non-linear exponential regression models. Based on the comparative results, the proposed model had higher prediction performance than the other three traditional models which has been widely used in the literature. To conclude, the proposed model could be used as a robust approach to handle uncertainty and complexity existed in accident data. In general, ANFIS model can be an effective tool for transportation agencies since intervention decisions and plans aiming at improving road safety depend on the prediction capabilities of a system.