Automatic identification method for driving risk status based on multi-sensor data

Real risk status detection is an effective way to reflect risky or dangerous driving behaviors and therefore to prevent road traffic accidents. However, a driver’s risk status is not only difficult to define but also uncontrollable and uncertain. In this study, a simulated experiment with 30 drivers was conducted using a driving simulator to collect the multi-sensor data of road conditions, humans, and vehicles. The driving risk status was classified into three states (0 - incident, 1 - near crash, or 2 - crash) on the basis of the playback system of the driving simulator. The experimental data were pre-processed using the cubic spline interpolation method and the time-windows theory. A driving risk status identification model was established using the C5.0 decision tree algorithm, and the receiver operating characteristic curve (ROC) was adopted to evaluate the performance of the identification model. The results indicated that respiration (RESP), vehicle speed (SPE), SM_FATIGUE, distance to the left lane (LLD), course angle (CA), and skin conductivity (SC) had a significant correlation (p < 0.05) with the driving risk status. The identification accuracy of the C5.0 decision tree algorithm was 78%, and the areas under the ROC were 0.934, 0.77, and 0.845, respectively. Moreover, compared with other four identification algorithms, the algorithm performance evaluation indexes TPR (0.780), precision (0.753), recall (0.78), F-measure (0.756), and kappa (0.884) of the C5.0 decision tree were all the best. The conclusion can provide reference evidence for danger warning systems and intelligent vehicle design.