Toward Personalized Car-Following Behaviors from Driver Data Using Learned and Hybrid Control Strategies

This paper introduces a method to reproduce personalized car-following behaviors from an expert human driver. With the basis of model predictive control (MPC) in the car-following control strategy, a data-driven model with deep neural network architecture is combined to perform the car-following style of a specific driver. By using the imitation learning method, the network model can directly map raw input from sensors to decision-making information in terms of speed tendencies in a human-like manner. Further, the hybrid strategies are designed to integrate the learned network model with the MPC controller. With the learned car-following behaviors from the expert driver, potential sources of discomfort such as jerk and uncomfortable speed changes can be effectively limited while maintaining safety in car-following maneuvers. The proposed control framework is implemented on an edge-computing platform and evaluated through real-vehicle experiments to demonstrate the capabilities to run learned network models. The promising results show similar car-following behaviors of the proposed control system to the expert driver and investigate the efficiency compared to ACC of commercial vehicles.