Real-Time Rear-End Conflict Prediction and Its Causal Analysis Using Interaction Trajectory Data

Most of the existing research on traffic conflict prediction currently adopts macroscopic aggregated traffic flow state features and road environment features for prediction modeling, which rarely considers the interaction behavior between related vehicles. So, we utilized microscopic interaction trajectory data of related vehicles for conflict prediction. We employ Gated Recurrent Unit to learn and encode a large amount of conflict and non-conflict trajectory data and perform binary classification, and it can accurately predict 93% of conflicts 2 s in advance. The Local Interpretable Model-agnostic Explanations tool is used to explain the causal relationships between results and input features. The results show that the high speed of the following car and the low speed of the leading car during the interaction process can promote conflicts. Moreover, the conflicts are dynamic behaviors between vehicles, which will evolve into conflicts when certain factors at a specific time point meet certain conditions.