关键词: 城市交通, 事故风险预测模型, 随机森林, 越江越海隧道入口段, 追尾事故

Abstract: This paper proposes a crash risk prediction model to evaluate the rear-end accident risk on the entrance section of cross-river and cross-sea tunnels. The vehicle speed data is extracted from the roadside monitoring videos at the entrance of the Shanghai Yangtze River Tunnel Bridge. The actual traffic flow conditions are obtained through the clustering algorithm. The average speed is used as the basis for the experimental parameters. The driving simulation scenarios involve different traffic flow conditions including congestion, close to congestion, and free flow condition. The weather conditions include sunny, rainy, and snowy. A high-risk situation of rear-end collision is included in each simulation scenario to analyze the driver's emergency response behavior. The correlation analysis and random forest method are used for variable screening and importance sorting. It was found that the number of vehicle operating states such as the head distance, speed difference, and acceleration standard deviation have significant impact on the occurrence of accident. The random oversampling method is used to improve the random forest model. The results indicate that among all the combinations, the short-term rear-end collision risk prediction model based on the random oversampling- random forest algorithm shows the best results. The Area Under Curve (AUC) index of the modified random forest model is increased by 6.8% compared to the traditional random forest model

Key words: urban traffic, crash risk prediction model, random forest model, entrance section of cross-river and crosssea tunnels, rear-end accident