A proactive crash risk prediction framework for lane-changing behavior incorporating individual driving styles

Area of Science:

• Traffic Safety and Intelligent Transportation Systems
• Machine Learning Applications in Automotive Engineering

Background:

• Driving style significantly impacts traffic safety, yet its interaction with lane-changing risk is poorly understood.
• Current advanced driver-assistance systems (ADASs) lack personalized risk prediction capabilities for lane-changing maneuvers.
• Individual driving styles necessitate tailored risk assessment for effective driver support.

Purpose of the Study:

• To propose a personalized lane-changing risk prediction framework integrating individual driving styles.
• To identify and analyze the driving styles and their associated risk factors during lane-changing.
• To enhance the safety of lane-changing decisions through personalized risk information services.

Main Methods:

• Development of driving volatility indices based on vehicle interactive features.
• Application of a dynamic clustering method to identify driving styles and optimal time windows.
• Utilizing Light Gradient Boosting Machine (LightGBM) with Shapley additive explanation for risk prediction and factor analysis.
• Evaluation using the highD trajectory dataset.

Main Results:

• Spectral clustering with a 3-second window accurately identifies driving styles during lane-changing intention.
• LightGBM demonstrates superior performance in personalized lane-changing risk prediction compared to other machine learning methods.
• Aggressive drivers exhibit higher lane-changing risk due to seeking greater freedom and neglecting surrounding vehicle states.

Conclusions:

• The proposed framework effectively predicts personalized lane-changing risk by considering driving styles.
• Driving style identification and risk analysis provide crucial insights for ADAS development.
• Findings support the creation of personalized warning systems for safer lane-changing maneuvers.