Stacking-based ensemble learning method for cognitive distraction state recognition for drivers in traditional and connected environments

Compared with the traditional environment, a connected environment in which multiple types of intelligent vehicles share roads will be more complicated. However, do changes in the traffic environment affect the driver's distraction level and distraction recognition? Relatively few studies have addressed these questions. This study constructs traditional and connected environments and conducts driving simulation experiments. Distracted driving data for 60 drivers were collected in each environment, and the one-way repeated analysis of variance (ANOVA) method was used to select the distraction indices with significant differences as the input indices. An independent t-test was applied to analyse changes in the distraction indices in the two environments. The results show that the distraction indices, including the standard deviation (SD) of the horizontal fixation angles, blink frequency, mean pupil diameter, mean saccade speed, SD of the steering wheel angle, mean speed and SD of the lane acceleration, differ in the two environments; therefore, the distraction recognition model in the traditional environment is no longer suitable for connected environments. A distraction state recognition model developed with a stacking-based ensemble learning algorithm composed of a support vector machine (SVM), extreme gradient boosting (XGBoost), long short-term memory (LSTM), and bi-directional LSTM with an attention mechanism (AT-Bi-LSTM) was proposed for each environment. The recognition accuracy reaches 95.71% in the traditional environment, and the accuracy rate reaches 97.92% in the connected environment. Compared with traditional algorithms, the proposed approach yields greater recognition accuracy and can be applied in intelligent vehicle distraction early warning assistance systems to improve road safety.