DCFS-based deep learning supervisory control for modeling lane keeping of expert drivers

Abstract

In this paper, a novel driver model for lane keeping is proposed to replicate the steering behavior of expert drivers. Specifically, a feedforward-feedback control scheme mocking expert drivers is adopted: the feedforward controller plays a leading role, which is a data-driven model based on deep convolutional fuzzy systems (DCFS), and for the sake of human-simulation and guaranteed stability, a supervisory feedback controller is designed, which works only if the state hits the set boundary. Comparing with the previous driver models, the key novelty of the paper is to introduce the “motor intermittency” of human behavior into driver modeling, which is an important issue for biological modeling of the drivers. Simulations on the joint platform of PreScan and CarSim show that the newly presented driver model has better matching performance to the expert drivers comparing with the two different types of advanced model predictive control (MPC) controllers. The proposed driver model has the potential application for semi-automated vehicles to provide human-like qualities for automated driving, which may be one of the essential points to promote the comfort when the driver hands over the steering authority, and improve the transition smoothness in the scenario of human vehicle co-piloting.