Lane-change risk prediction model based on traffic context and driving styles

Area of Science:

• Road safety
• Artificial intelligence in transportation
• Vehicle dynamics

Background:

• Existing lane-change risk models lack adaptability in complex traffic.
• Traffic context and driving styles are underutilized factors in risk prediction.

Purpose of the Study:

• To develop an integrated lane-change risk prediction model.
• To enhance driving safety by considering traffic context and driving styles.
• To improve the accuracy and adaptability of lane-change risk assessment.

Main Methods:

• Attention-LSTM for traffic context and lane change type identification.
• LSTM-Stacked Denoising Autoencoder (LSTM-SDAE) for driving style characterization.
• Light Gradient Boosting Machine (LGBM) for risk prediction, utilizing Shapley additive explanations (SHAP) for feature importance.

Main Results:

• Model accuracy improved from 88.79% to 95.05% with traffic context integration.
• The LGBM model with integrated features outperformed other algorithms.
• Feature importance varied for left vs. right lane changes, highlighting lateral and longitudinal velocities and driving style.

Conclusions:

• Integrating traffic context and driving styles significantly enhances lane-change risk prediction accuracy.
• The proposed model offers valuable insights for improving driving safety systems.
• This approach provides a foundation for more robust decision-support systems in complex traffic scenarios.