Modeling lane-changing spatiotemporal features based on the driving behavior generation mechanism of human drivers

Dongjian Song

Abstract: Accurate lane-changing (LC) modeling is important to realizing human-like LC in intelligent vehicles (IVs). The LC behavior of human drivers is often determined by LC spatiotemporal features such as the target lane, LC starting point, duration, target position and velocity. Therefore, to accurately reproduce the LC behaviors of drivers, this paper proposes an LC spatiotemporal feature model (LSFM) based on the driving behavior generation mechanism. First, we consider the generation of LC behaviors of human drivers as a Markov decision process and establish the framework of the LSFM by semantically deconstructing the LC behavior generation mechanism. Then, the cognitive and behavioral characteristics of drivers are described through human-like reward functions. Furthermore, the selection of actions of the LSFM is converted to the selection of LC spatiotemporal features by establishing the expected trajectory space according to spatiotemporal features. The expected trajectory space is resized and pruned based on statistics and safety constraints. Thus, the sampling efficiency and safety of the LSFM are improved. Finally, the human-like reward function weights are recovered from High D by maximum entropy inverse reinforcement learning. In addition, the LSFM divides LC into anticipation and relaxation, for which it designs different reward functions and expected trajectory spaces, which further improves the modeling accuracy. The verification results on the naturalistic driving data show that the LSFM can more accurately model LC spatiotemporal features than current models, and it has good generalizability to provide important support for human-like LC in IVs.