Abstract:

The freeway merging area is a high-risk zone for traffic accidents due to frequent lane changes, complex driving conditions, and significant traffic conflicts. Based on the German exiD vehicle trajectory data, this study categorizes vehicle merging patterns in freeway merging zones and proposes a risk characterization index system, using the Time-to-Collision (TTC) theory to evaluate the risk at the moment of merging and throughout the merging process. Merging risk prediction models are developed using machine learning techniques, including XGBoost, LightGBM, and GBDT. Additionally, the SHAP theory is applied to analyze the causal factors contributing to merging risks in freeway merging areas. Experimental results show that the proposed merging risk prediction model achieves an overall accuracy of 95.52%, outperforming models such as Random Forest, LightGBM, and GBDT in terms of accuracy, precision, recall, and F1-score. The model comparison results indicate that risk recognition models incorporating factors like merging duration and urgency achieve higher accuracy. Furthermore, the merging risk in the freeway merging area is closely related to factors such as the average speed difference between the merging vehicle and the preceding vehicle, merging duration, and the maximum speed difference with the preceding vehicle.

Key words: traffic safety, freeway merging areas, traffic conflicts, risk identification, machine learning