Role of Risk Perception in Predicting the Intentions of Wakeful and Drowsy Drivers

Driver intention recognition algorithms play a crucial role in understanding driver–vehicle interactions. In highly automated vehicles, drivers are occasionally instructed to take control. Driver–vehicle interaction is, therefore, an important research topic. This paper uses risk levels and risk dynamics to provide a ground for developing intention recognition algorithms. It also investigates the effect of drowsiness on a driver’s ability to perform defensive actions. As human controllers are primarily event-based, the driver is modeled as an event-based controller that monitors risk-based signals and interacts with the vehicle pedals. A Gaussian support vector machine (SVM) model was used to predict the reaction type. A car-following experiment was conducted, in wakeful and drowsy phases, to assess wakeful driver actions and the ability of drowsy drivers to confine risk-based variables. Twenty-one drivers were tested in a wakeful state. The results achieved maximum and minimum accuracies of 81.9% and 60.8% in predicting the type of driver’s defensive action. The rate of the accelerator pedal release was affected by the risk and its time derivative, while the duration of switching from the accelerator to the brake was only affected by the time derivative of risk. Subsequently, the effects of drowsiness on driver performance were explored. Finally, a long short-term memory (LSTM) model was employed to predict drivers’ actions. The findings suggest that monitoring drivers’ risk estimations and their defensive pedal responses reveal patterns for predicting both the timing and the type of driver reactions.