An integrated framework for driving risk evaluation that combines lane-changing detection and an attention-based...

Area of Science:

• Artificial Intelligence
• Road Safety Engineering
• Machine Learning

Background:

• Rising traffic accidents pose a significant threat to public safety.
• Predicting and mitigating risky driving scenarios is crucial for enhancing road safety.

Purpose of the Study:

• To develop and validate a predictive model for identifying risky driving scenarios.
• To improve road safety through real-time risk prediction and driver guidance.

Main Methods:

• A framework combining an integrated lane-changing detection model and an attention-based long short-term memory (LSTM) prediction model was developed.
• Machine learning methods were evaluated for lane line detection, selecting an ultrafast network.
• Attention-based LSTM was compared against standard LSTM for prediction accuracy, recall, precision, and F1 score.
• Shapley additive explanation (SHAP) analysis was employed for model interpretability.

Main Results:

• The ultrafast lane detection network demonstrated high efficiency with a running time of 4.1 ms and a speed of 131 fps.
• The attention-based LSTM model achieved high prediction performance with 96% accuracy, 98% precision, 96% recall, and 97% F1 score.
• SHAP analysis revealed that lane changes, yaw rate, speed stability, vehicle speed, and acceleration are key factors influencing driving risk.

Conclusions:

• The attention-based LSTM model significantly outperforms the standard LSTM in convergence speed and prediction accuracy for identifying risky driving scenarios.
• Lane change behavior was identified as the most critical factor impacting driving risk.
• The developed model offers potential for real-time risky scenario prediction, driver warnings, and informed decision-making to enhance road safety.