A Traffic Flow Simulation Framework for Learning Driver Heterogeneity from Naturalistic Driving Data using Autoencoders

Abstract

This paper proposes a novel data-centric framework for microscopic traffic flow simulation with intra and inter driver heterogeneity. We utilized a naturalistic driving corpus of 46 different drivers to learn and model the behavior divergence of Japanese drivers. First, ego-driver behavior signals are used to extract unique features of each driver with an auto-encoder. Then, using these features, drivers are divided into groups using unsupervised clustering algorithms. For each driver group, a feedforward neural network is trained for predicting the desired speed given the road topology. The trained network is then used in a microscopic traffic flow model for simulations. We used a macroscopic traffic survey conducted in Japan to evaluate the proposed framework. Our findings indicate that the proposed framework can simulate a realistic traffic flow with high driver heterogeneity.