Real-time driving risk prediction using a self-attention-based bidirectional long short-term memory network based on...

Area of Science:

• Artificial Intelligence
• Transportation Safety
• Machine Learning

Background:

• Predicting driving risks is crucial for collision prevention.
• Existing methods often rely on single data sources and lack advanced modeling or time window analysis.

Purpose of the Study:

• To propose a novel Att-Bi-LSTM model for multi-source driving risk prediction.
• To evaluate the model's performance against established machine learning algorithms.

Main Methods:

• Collected multi-source driver data (demographic, operational, visual, physiological, kinematic) from simulation tests.
• Developed an Att-Bi-LSTM network and compared it with CNN, CNN-LSTM, CatBoost, LightGBM, and XGBoost.
• Utilized observation, interval, and prediction time windows for model input/output generation.

Main Results:

• The Att-Bi-LSTM model achieved a macro-average F1-score of 0.914, outperforming all comparison models.
• Ablation studies confirmed the effectiveness of Bi-LSTM layers and self-attention mechanisms.
• Multi-source data improved the F1-score by 0.061 compared to using only kinematic data.

Conclusions:

• The proposed Att-Bi-LSTM model offers an effective approach for driving risk prediction.
• Model performance is sensitive to time window configurations, with optimal observation windows improving accuracy.
• This research supports the development of enhanced advanced driver assistance systems.