Identification of Accident Causal Factors and Prediction of Hazardousness of Intersection Approaches

Abstract

Safety improvements were evaluated at signalized intersection approaches, where about 30% of road crashes occur in Singapore (in Singapore, vehicles are driven on the left side of the road). Identification of accident causal factors and prediction of hazardousness are essential tasks before safety treatments can be prescribed and prioritized. Statistical methods were used as a novel method for the identification of causal factors affecting accident frequencies, while nonparametric regression techniques were used for the prediction of intersection approach hazardousness. The statistical methodology used, the random-effect negative binomial model, indicates that uncontrolled left-turn slip roads, sight distances of less than 100 m, large numbers of signal phases, the use of permissive right-turn phases, the existence of horizontal curves, and total and left-turn volumes may increase the likelihood of accident occurrence. By nonparametric regression, the subtractive clustering methodology in fuzzy logic was used to derive an efficient and effective data set from the original data, which exhibited a high level of noise. This clustered data set was used to predict hazardousness. The prediction model, a generalized regression neural network, showed reasonable prediction accuracy. The model was capable of predicting hazardousness of about 65% for the total production data set, with a difference between the actual and the predicted values of less than 0.1. Moreover, the results suggest that the nonparametric regression network works well in multidimensional measurement spaces, in which hazardousness is assumed to be a function of several geometric, traffic, and traffic control measures.