A high-resolution trajectory data driven approach for vehicle conflict risk prediction considering driving style

Frequent traffic collisions result in considerable loss of life and property worldwide. To improve traffic safety, it is essential to explore the relationship between real-time vehicle characteristics and crash risk in car-following scenarios and to devise effective countermeasures. This study presents a personalized car-following risk prediction framework that accounts for individual driving styles. By integrating vehicle-specific and interactional features with conflict data, internal relationships are analyzed using high-resolution trajectory information of the HighD dataset from Germany, which provides the microscopic motion states of vehicles. Conflict events are detected through surrogate safety measures based on a Time-to-Collision (TTC) threshold of less than 4 seconds. A binary logistic regression model is applied to quantify the impacts of vehicle features on conflict risk. Additionally, driving styles are categorized into cautious, normal, and aggressive groups using a K-means clustering algorithm. Based on these classifications, four machine learning models—Decision Tree (DT), Random Forest (RF), Adaptive Boosting (AdaBoost), and Light Gradient Boosting Machine (LightGBM)—are employed to predict car-following risks. Among these models, LightGBM demonstrates the highest prediction accuracy (0.98). The results suggest that the proposed framework effectively estimates car-following conflict risks, offering valuable insights for active traffic management and safety intervention strategies.