Predict the risk feeling for drivers of autonomous cars: an application of deep learning methods

Simulation is used to assess safety provided by autonomous vehicle algorithms. However, safety derived by computation systems can have significant gaps with driver’s feeling of safety. Thus, to improve validation by simulation tools, autonomous vehicle designers need to implement criteria for risk perception assessment. We demonstrate, in this study, that risk feeling is significantly related to some personal characteristics of the driver and to his past and current driving events. We propose to compare three deep learning-based networks to model it. The outcome of this cognitive driver model is a classification on 5 risk levels felt by the driver. Two metrics are adopted as the measure of the models’ accuracy: the area under the curve and the F1-score. They show accurate prediction of the driver emotional state in autonomous driving scenarios of car-following and overtaking maneuvers, which corresponds to most highway situations. The main improvement factor of this method is the integration of individual driver characteristics in the learning model. Thus, simulation enables further design of a secure automatic driving system as well as the design of an automatic driving behavior fitted for the driver cluster targeted.