Development and Evaluation of Two Learning-Based Personalized Driver Models for Car-Following Behaviors

摘要

Personalized driver models play a key role in the development of advanceddriver assistance systems and automated driving systems. Traditionally,physical-based driver models with fixed structures usually lack the flexibilityto describe the uncertainties and high non-linearity of driver behaviors. Inthis paper, two kinds of learning-based car-following personalized drivermodels were developed using naturalistic driving data collected from theUniversity of Michigan Safety Pilot Model Deployment program. One model isdeveloped by combining the Gaussian Mixture Model (GMM) and the Hidden MarkovModel (HMM), and the other one is developed by combining the Gaussian MixtureModel (GMM) and Probability Density Functions (PDF). Fitting results betweenthe two approaches were analyzed with different model inputs and numbers of GMMcomponents. Statistical analyses show that both models provide good performanceof fitting while the GMM--PDF approach shows a higher potential to increase themodel accuracy given a higher dimension of training data.