Car-Following Model Based on Deep Learning and Markov Theory

Abstract

A car-following (CF) model can reproduce various micro traffic phenomena and plays a crucial role in traffic theory. In this study, we combine Markov theory and a gated recurrent unit (GRU) neural network (NN) to propose a new CF model. Next-generation simulation (NGSIM) data were used to generate the Markov chain and train the GRU-NN. Considering the memory effects, we predicted each vehicle’s state at the next time step by the headways and speeds in the last several time steps. Simulations were used to test the merits of the proposed CF model under some given scenarios. The results indicate that the proposed CF model has high accuracy and can enhance the stability of trajectory prediction in simulation, which provides a new approach for micro traffic simulation.