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

Early warning of driving risks can effectively prevent collisions. However, numerous studies that predicted driving risks have suffered from the use of single data sources, insufficiently advanced models, and lack of time window analysis. To address these issues, this paper proposes a self-attention-based bidirectional long short-term memory (Att-Bi-LSTM) network model to predict driving risk based on multi-source data. First, driving simulation tests are conducted. Driver demographic, operation, visual, and physiological data as well as kinematic data are collected. Then, the driving risks are classified into no risk, low risk, medium risk, and high risk. Next, the Att-Bi-LSTM model is constructed, and convolutional neural network (CNN), CNN-LSTM, CatBoost, LightGBM, and XGBoost are employed for comparison. To generate the inputs and outputs of the models, observation, interval, and prediction time windows are introduced. The results show that the Att-Bi-LSTM model using early-fusion method significantly outperforms the five comparison models, with a macro-average F1-score of 0.914. The results of ablation studies indicate that the Bi-LSTM layers and self-attention layer have achieved the expected effect, which is crucial for improving the model's performance. As the interval or prediction time window is extended, the accuracy of the prediction results gradually decreases. However, as the observation time window is extended, the results first improve and then become stable. Compared to using only relative kinematic data, using all data (i.e., multi-source data) is shown to improve the F1-score by 0.061. This study provides an effective method for driving risk prediction and supports the improvement of advanced driver assistance systems.