Incorporation of Human Factors to a Data-Driven Car-Following Model

Abstract

In recent years, there have been intensive efforts to consider human factors (HFs) in the modeling of human driver behavior. In particular, “engineering” car-following models widely used in traffic simulation have been extended to include HFs. This extension is needed to generate critical situations, which are often attributable to human error. However, incorporation of reaction processes requires advanced models that take driver predictions and delayed responses into account. In this paper, a methodology for integrating HFs into driver behavior modeling is developed based on a long short-term memory architecture. The proposed methodology employed a three-layer psychological concept: perception, information processing, and action. The perception layer modeled (imperfect) estimation of visually received stimuli. Information processing included short-term memory and the projection of perceived stimuli into the near future. The executed action, based on the sensed as well as anticipated dynamic driving state, was delayed by the perception–reaction time. To represent individual differences among driver types, the available training dataset was classified in four clusters according to observable driver characteristics. The methodology was demonstrated with data recorded at an urban signalized intersection. Model performance was compared with that of two established engineering models, the intelligent driver model and the (extended) full velocity difference model. The results indicated that the human driver model developed here showed superior performance in replicating real-world trajectories compared with existing models and was able to represent the varying driving strategies of different groups.