Prediction of Freeway Rear-End Collision Risk Based on Trajectory Data

This paper presents a method to predict the instantaneous rear-end conflict risk of individual cars based on trajectory data that uses Time to Collision (TTC) to identify the risk. Features from 1 to 5 s before the risk occurs are extracted, including the vehicle’s driving status, interaction with the preceding car and macro traffic state, while the vehicles without risk are extracted as control. Filtered features are used to build Random Forest and XGBoost models to predict the rear-end conflict risk from 1 to 5 s before. The results show that the instantaneous conflict prediction models are effective, with accuracy and recall from 1 to 3 s prior to conflict surpassing 92% and 76%, respectively. The vehicle’s speed, speed difference with the preceding vehicle, space and time headway, as well as the average speed and average space headway of the lane are all important influencing factors.