A personalized driving risk assessment and rolling prediction method by integrating multiple indicators

There are significant differences in driving styles among different drivers. A personalized perspective on driving risk assessment and prediction can help to proactively prevent traffic safety and provide a risk prediction basis for intelligent driver assistance systems. A personalized driving risk assessment and rolling prediction method is proposed to predict a driver's potential accident risk in real-time by integrating vehicle dynamics and Heart Rate Variability (HRV) indicators. A natural driving experiment is designed to obtain Critical Incident Events (CIEs) and changes in each indicator. The frequency of CIEs is used as a driving risk characterization. To identify the significant personalized indicators affecting various CIEs, an improved Bayesian network model is developed to obtain the influence mechanism of each indicator on CIEs. The Dynamic Time Warping Barycenter Averaging (DBA) method is used to calculate the representative series of each significant indicator, which can obtain the characteristic time series under different risk levels. The weights of each CIE class are calculated by the entropy weight Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) to get the risk score calculation rule. These scores are then clustered by the Fuzzy C-means (FCM) algorithm to determine the different risk levels. Finally, the Bayesian Optimization (BO) -based Bidirectional Gated Recurrent Unit (BiGRU) is integrated with a Convolutional Neural Network (CNN) and Extended Kalman Filter (EKF) to construct the BCBGE (BO-CNN-BiGRU-EKF) model, which enables continuous prediction of driving risk. Results from a natural driving experiment involving 60 drivers in Xi'an indicate that driving risk can be grouped into four levels. A personalized risk indicator analysis was conducted for each driver. The results indicate that each type of CIE is associated with three to four key indicators of vehicle dynamics or HRV. When the observation window length is 3.8 s and the prediction window length is 2.4 s, the proposed rolling prediction model achieves an accuracy of 92.03%, which is 3.37% to 15.88% higher than the accuracies obtained using the GRU, Bidirectional Long Short-Term Memory (Bi-LSTM) -EKF, and BO-CNN-BiLSTM-EKF models.