军事知识服务平台
融入驾驶风格的SAC跟驰模型
Soft Actor-Critic based car-following model integrated with driving style
摘要 [目标]为提升自动驾驶车辆跟驰行驶的安全性、舒适性与场景适配性,解决传统车辆跟驰模型无法适配差异化驾驶风格、难以满足智能驾驶多样化行驶需求的问题,构建一个可自适应不同驾驶行为特征的智能车辆跟驰模型.[方法]依托 SAC 强化学习算法样本利用率高、参数调试简便、适配连续动作控制的技术优势,引入驾驶风格因子完成算法优化改进.基于中国驾驶行为数据集 CQSkyEyeX,通过降维聚类方法挖掘、划分差异化驾驶风格类别,构建贴合不同驾驶特征的专属奖励函数,将其嵌入 SAC 算法框架,搭建自适应驾驶风格的车辆跟驰模型.依托 SUMO仿真平台实现模型在线更新与动态仿真验证,并与传统智能驾驶员模型开展多维度性能对比试验.[结果]仿真试验结果表明,优化后的 SAC 跟驰模型可有效规避车辆碰撞风险,全程保障跟驰行驶安全性.相较于传统智能驾驶员模型,该模型能够高度还原真实道路驾驶轨迹,轨迹拟合精度显著提升,可精准区分、适配激进、稳健、保守等各类差异化驾驶风格.[结论]融入驾驶风格因子的改进 SAC 跟驰模型具备优异的环境自适应能力与驾驶行为拟合能力,突破了传统跟驰模型驾驶风格同质化的局限,可有效适配复杂道路交通场景下的多样化智能驾驶需求,为自动驾驶车辆跟驰控制算法的优化与工程应用提供技术参考.
Abstract:
[Objective]Safety,comfort and scenario-adaptability are essential for car-following driving of autonomous vehicles.Conventional car-following models fail to adapt to differentiated driving styles and satisfy diverse intelligent driving demands.This study proposes a car-following model adaptive to diverse driving behavioral characteristics.[Method]Soft Actor-Critic(SAC)reinforcement learning algorithm featured high sample efficiency,simple parameter calibration and favorable performance in continuous action control.Driving style factors were introduced to optimize SAC algorithm.The dimensionality reduction and clustering methods were applied to classify differentiated driving styles,based on the dataset CQSkyEyeX.Targeted reward functions were established and embedded into SAC framework to construct the adaptive car-following model.The model updating and dynamic simulation verification were online conducted by using SUMO.The comparative test on multi-dimensional performance was conducted between the proposed model and traditional intelligent driver model.[Result]The optimized SAC model avoids vehicle collision and guarantees continuous car-following safety.Compared with traditional intelligent driver model,the optimized model reconstructs real-road driving trajectories with high accuracy and improves trajectory fitting performance.The model effectively identifies and adapts to aggressive,moderate and conservative driving styles.[Conclusion]The improved SAC model embedded with driving style factors presents outstanding environmental adaptability and driving behavior fitting capacity.It eliminates the driving style homogeneity limitation of conventional car-following models,satisfies diverse intelligent driving demands in complex traffic scenarios,and provides a technical reference for the optimization and engineering application of autonomous vehicle car-following control algorithms.
刊 名 公路交通科技 ISTICPKU
年,卷(期): 2026, 43(9)
分类号: U461.91
机标分类号: U495 TP391.9 TP273.1
在线出版日期: 2026年10月22日
基金项目: 51975428:国家自然科学基金,