A Learning-Based Personalized Driver Model Using Bounded Generalized Gaussian Mixture Models

Abstract

Individual driver's driving behavior plays a pivotal role in personalized driver assistance systems. Gaussian mixture models (GMM) have been widely used to fit driving data, but unsuitable for capturing the data with a long-tailed distribution. Though the generalized GMM (GGMM) could overcome this fitting issue to some extent, it still cannot handle naturalistic data which is generally bounded. This paper presents a learning-based personalized driver model that can handle non-Gaussian and bounded naturalistic driving data. To this end, we develop a BGGMM-HMM framework to model driver behavior by integrating a hidden Markov model (HMM) in a bounded GGMM (BGGMM), which synthetically includes GMM and GGMM as special cases. Further, we design an associated iterative learning algorithm to estimate the model parameters. Naturalistic car-following driving data from eight drivers are used to demonstrate the effectiveness of BGGMM-HMM. Experimental results show that the personalized driver model of BGGMM-HMM that leverages the non-Gaussian and bounded support of driving data can improve model accuracy from 23 $ \sim$ 30% over traditional GMM-based models.