Human-driving like lane-changing behavior of autonomous vehicles based on asymmetric risk field and reinforcement learning

Abstract

Lane-changing (LC) behavior is a common and safety-risky behavior in traffic system, accurately quantizing the risk during LC process and establishing a reasonable LC model are crucial for autonomous vehicles to complete LC process like human-driving vehicles. So far, the risk in LC process is mainly assumed to be symmetric in existing studies and the different safety risks posed by different types of vehicles are ignored. To explore the safety risks posed by different types of vehicles in real traffic, an asymmetric risk field LC model from the perspective of asymmetric risk is established in this paper. In this model, the asymmetric risk is calculated in view of the vehicle size, and the vehicle size is presented as volume based on natural dataset. Also, the risk threshold that is introduced to depict the LC behavior in line with human-driving characteristics is calibrated by applying reinforcement learning (RL) method and NGSIM dataset. Finally, comparison simulation between the proposed model and the symmetric risk model is carried out and the simulation results illustrate that the longitudinal error (LE), the mixed gap error (MGE), and the model error (ME) of the proposed model with real data is lower than that of the symmetric risk model with real data. It shows that the proposed model is more consistent with the real LC trajectory than the symmetric risk model.