Key variables identification and proactive assessment of real-time traffic flow accident risk on urban expressway

Abstract

Based on accident data and detector data collected on two expressways in Shanghai, important variables for model construction were selected from the data of traffic flow within 5~10 min before the accident with random forest model. Then, the Bayesian network (BN) model based on the Gaussian mixture model and expected maximum algorithm was established for the analysis of real-time traffic flow state and accident risk. Meanwhile, the transferability of BN model was also assessed. The results show that BN model built with selected important variables is better than that with direct detection data, with the accident prediction accuracy rate of 82.78%. The results of the transferability show that the improved BN model is still better than the traditional model, though the accident prediction accuracy of BN model decreases.