考虑驾驶员特性的个性化跟驰控制策略研究

Study of Personalized Car-following Control Strategy by Considering the Driver Characteristics

为提高车辆自主跟驰功能的个性化程度,使之能适应不同驾驶员的驾驶风格,提高乘员对自动驾驶功能的接受度,提出了一种基于深度强化学习的个性化跟驰控制策略.首先基于模拟驾驶试验平台进行驾驶员在环试验,获取真实驾驶员跟驰数据.根据车辆跟驰动力学,建立了连续动作空间决策模型.构建了基于Actor-Critic的深度强化学习架构,并综合考虑跟驰过程的安全性、舒适性和宜人性设计了奖励函数,通过双延迟深度确定性策略梯度(Twin Delayed Deep Deterministic Policy Gradient,TD3)对决策模型进行训练.基于CARLA模拟器的仿真结果表明,本研究提出的个性化跟驰控制策略在保证车辆自主跟驰过程稳定性和安全性的前提下,其决策结果更接近驾驶员驾驶习性.

In order to improve the degree of personalization of the car-following control which makes it can adapt to the driving styles of different drivers,a personalized car-following control algorithm based on deep reinforcement learning is proposed in this paper.Firstly,the driving simulation platform was established to adopt the driver-in-loop experiment for obtaining the real driver following data.Then,based on the car-following dynamics,the continues action space decision model was built.The deep reinforcement learning model was introduced with actor-critic architecture and the reward function was designed considering the safety,comfort,and driver behavior.The Twin Delayed Deep Deterministic Policy Gradient(TD3)algorithm was proposed to train the decision model.Finally,the results of simulation based on the CARLA simulator demonstrated that the personalized car-following control strategy proposed in this paper is closer to the driving habits of human drivers on the premise of ensuring the stability and safety of the autonomous car-following process.