文章摘要
基于深度强化学习的自动驾驶分段决策方法
Subsection Decision-making Method for Autonomous Driving Based on Deep Reinforcement Learning
投稿时间:2023-11-27  修订日期:2024-01-02
DOI:
中文关键词: 自动驾驶  驾驶决策  道路分段  小波分解  深度强化学习
英文关键词: autonomous driving  driving decision  road segmentation  deep reinforcement learning  wavelet transform
基金项目:国家自然科学基金项目(面上项目,重点项目,重大项目)
作者单位
王春淇 大连理工大学 机械工程学院 
张明恒* 大连理工大学 机械工程学院 
周俊平 大连理工大学 机械工程学院 
姚宝珍 大连理工大学 机械工程学院 
石佳伟 大连理工大学 机械工程学院 
摘要点击次数: 70
全文下载次数: 0
中文摘要:
      目前,自动驾驶已成为车辆工程技术领域研发的热点,安全、舒适、高效是自动驾驶决策系统开发需要考虑的基本性能要求。本文基于道路线形分段,在不同道路分段采用差异化的车辆控制策略,以满足自动驾驶决策系统的性能要求。首先,鉴于道路线形特征的复杂变尺度特性,本文通过小波变换及最大类间方差法(Otsu Algorithm)将道路特征与变换系数进行映射并实现道路自适应分段。其次,为提高模型决策能力,基于深度强化学习(Deep Reinforcement Learning, DRL)与奖励分解架构将驾驶决策任务分解为横、纵向两个并联决策子任务并分别构建决策模型与奖励函数,并设计一种动作掩蔽策略来提升模型训练速度。最后,进行一系列实验来验证了本文所提出模型的有效性,实验结果表明,与传统的DRL算法相比,引入奖励分解架构与动作掩蔽策略的DRL算法在保证驾驶系统可靠决策的同时,在通行效率、安全性等方面均有提升。
英文摘要:
      At present, autonomous driving has become a hotspot in the field of vehicle engineering, the performance of safety, comfort and efficiency need to be considered basically while developing decision system. Based on the road alignment segmentation, this paper adopted differentiated vehicle control strategies in different road segments to achieve the autonomous driving performance standards. Firstly, given the complex and varying scale characteristics of road alignment attributes, this paper utilized the wavelet transform and Otsu algorithm to map road features to the transform coefficient and segment the road adaptively. Secondly, to enhance the decision-making ability, this study decomposed driving decision-making into two parallel sub-tasks utilizing the deep reinforcement learning (DRL) algorithm and the reward decomposition architecture. Each sub-task constructed the decision model and reward function separately. Additionally, this study developed an action masking strategy to speed up the model training. Finally, in order to validate the enhancement of the proposed model, this paper designed a series of experiments. The results indicate that the model proposed in this paper enhances driving safety and traffic efficiency, while ensuring reliable decision-making, in comparison to the traditional DRL algorithm.
View Fulltext   查看/发表评论  下载PDF阅读器
关闭