基于深度强化学习的智能模型车云端决策方法研究

发布时间：2021-09-24 01:09

　　提高交通效率的常用方法是控制交通信号灯以确保交通畅通,然而由于车辆行为的不可控,实际效果有限。随着智能网联汽车技术的发展,交通系统的云端控制中心不仅可以控制交通信号灯,还有可能直接控制车辆。在这种以云端为控制中心的交通管理模式下,云端决策能力是决定交通系统效率的关键因素。云端决策算法将是未来智能交通系统的关键技术。由于云端决策研究涉及多车协同,使用真实车辆进行研究的难度和危险性都很高。因此本文将以智能模型车为载体重点研究基于深度强化学习的云端决策算法。本文的研究工作可以大致分为三大部分:首先,本文提出了融合视觉与UWB的室内定姿定位算法,解决了模拟车位姿信息的准确获取。目前已有的室内定位方法是基于相机的检测,缺点是对目标的光线和颜色过于敏感,导致检测定位不够稳定可靠。因此,本文研究了不依赖相机的无线定位方式UWB,构建了基站自适应选择的UWB定位系统,解决多基站时间难同步和定位精度不稳定的问题。在此基础上,本文进一步研究了实时定位信息与地图先验信息的融合定位方法,实现了多目标的位置检测与跟踪。本文基于UWB定位系统实现了良好的定位效果,然而无法获取模型车姿态信息。因此进一步研究了相机检... 

【文章来源】：清华大学北京市 211工程院校 985工程院校 教育部直属院校

【文章页数】：94 页

【学位级别】：硕士

【文章目录】：
摘要
Abstract
Chapter1 Introduction
    1.1 Background
    1.2 Related Work
    1.3 Problem Statement
    1.4 Thesis Outline
Chapter2 Theory of Deep Reinforcement Learning
    2.1 Reinforcement Learning
        2.1.1 Element of Reinforcement Learning
        2.1.2 Markov Decision Process
        2.1.3 Dynamic Programming
        2.1.4 Learning Method
    2.2 Neural Network
        2.2.1 Basics of Neural Network
        2.2.2 Neural Network in Reinforcement Learning(Deep Reinforcement Learning)
Chapter3 The localization based on Fusion of uwb and camera
    3.1 Overall Structure of Localization System
    3.2 UWB Localization
        3.2.1 Ranging Process
        3.2.2 False Ranging Detection
        3.2.3 Solving Trilateration Algorithm
    3.3 Camera Localization
        3.3.1 Fish Eye Calibration
        3.3.2 Camera Detection
    3.4 Sensor Fusion Between Camera,UWB and Map Information
Chapter4 Deep Reinforcement Learning Method
    4.1 Training Environment
        4.1.1 Map Drawing
        4.1.2 Dynamics Model
        4.1.3 Steering Control Model
        4.1.4 State Action Space
        4.1.5 Reward Function
        4.1.6 Termination Stage
    4.2 Reinforcement Learning Models
        4.2.1 Deep Q learning with Experience Replays
        4.2.2 Asynchronous Advantage Actor Critic(A3C)
    4.3 Training Results
        4.3.1 Deep Q learning Network
        4.3.2 Asynchronous Advantage Actor Critic(A3C)
Chapter5 Deep Reinforcement Learning Validation and Evaluation
    5.1 Experimental Setup
    5.2 Motion Control of Intelligent Vehicle
    5.3 Localization Result
    5.4 Validation Decision Making Experiment Result
        5.4.1 Validation of RL Based Decision Through Simulation Software
        5.4.2 Validation of RL based Decision Through Model Cars
Chapter6 Conclusion
References
Acknowledgement
RESUME



本文编号：3406794