城市环境下无人驾驶车辆决策系统研究
发布时间：2018-07-31 07:39
【摘要】：随着计算机科学和机器人技术的发展,无人驾驶车辆在军事、民用和科学研究等诸多方面得到了广泛的应用,它集中了结构学、电子学、控制论和人工智能等多学科的最新研究成果,具有广阔的应用前景。 对于无人驾驶车辆来说,智能决策是其关键组成部分,是研究的热点之一。在城区环境中,由于驾驶场景复杂多变,交通参与者的行为难以预测,无法采用一个标准、统一的决策模型进行描述。为了解决城区复杂环境下无人驾驶车辆的决策、规划问题,本文通过学习人类驾驶员复杂场景下的决策过程,提出了一种新的驾驶行为决策模型的建立方法,并在此基础上,完成了无人车的运动规划。具体的研究内容如下所示： 1)介绍了无人驾驶车辆的研究意义,了解了无人驾驶车辆的国内外研究现状,调研了移动机器人智能决策及运动规划方法,分析比较了国外无人驾驶车辆决策系统的实现方式,对城市典型交通状况进行了描述和总结。基于驾驶员的任务需求以及其对车辆的决策过程及行为方式,提出了无人驾驶车辆在城市道路中进行决策规划的关键问题,明确了决策系统的设计准则。随后介绍了“智能先锋Ⅱ”无人驾驶车辆平台的各个组成模块,阐述了平台的工作原理及协作方式。基于决策系统时间和空间上的多分辨率的特点,设计了三层模块化决策系统框架,满足决策系统的实时性、自适应性和鲁棒性要求。 2)针对不同的驾驶环境及人类驾驶员的驾驶行为特征,采用层次有限状态机的方法建立城区环境下无人驾驶车辆行为决策模块。对驾驶员的复杂行为进行抽象和分解,并把分解所得的原子行为作为状态机的底层状态集合。同时,基于人类驾驶员复杂场景下的决策过程,提出了一种基于多属性决策方法的驾驶行为决策模型,抽取行驶过程中驾驶员关注的相关属性,判断和评价并获取最终驾驶行为模式,使得行为决策模式符合人类驾驶员的思维过程,解决了城市复杂交通场景下无人驾驶车辆的类人决策问题。设计了一种基于层次分析法——熵权法的驾驶行为矩阵赋权方法,建立基于驾驶经验与客观数据的权重体系,减弱主观随意性对决策结果的干扰,并且减少因样本数据不足带来的熵值法不准确的问题。结合TOPSIS优选和灰色关联分析两种方法,构建一种新的驾驶行为灰色理想值逼近模型,进行决策评判,使得被选方案不仅在空间位置上与最优方案较为接近,同时,其形状也贴近于最优方案,保证了所选驾驶行为的最优性。 3)研究基于径向基函数神经网络的运动规划方法。首先对城区环境中运动规划算法的设计原则进行了分析,明确了运动规划的难点。针对非结构化道路特征不明显,环境不可预测的特点,提出了一种基于径向基函数(RBF)神经网络的运动规划方法,通过随机提取可行驶区域内的离散参考点,采用正则化网络对数据做逼近处理,并以一种带遗忘因子的单输出RBF网络学习方法——梯度下降法对网络参数进行学习。生成的轨迹能够对任意道路形状进行拟合,并且满足车辆运动特性的约束。此外由于RBF网络是一种局部逼近网络,具有学习速度快的优点,对于动态变化的环境能够快速响应,满足车辆规划系统的实时性要求。 最后,本文基于城区真实道路环境,在“智能先锋Ⅱ”无人驾驶车辆平台上进行了实验,结果验证了决策系统设计方法的正确性和有效性。
[Abstract]:With the development of computer science and robotics, unmanned vehicles have been widely used in many fields, such as military, civil and scientific research. It focuses on the latest research achievements in many disciplines, such as structure, electronics, cybernetics and artificial intelligence, and has a broad application prospect.
Intelligent decision making is one of the key components of unmanned vehicle, which is one of the key parts of research. In urban environment, because of the complex and changeable driving scene, the behavior of the traffic participants is difficult to predict. It is impossible to use a standard and unified decision model to describe the unmanned vehicle in the complex environment of urban area. On the basis of learning the decision-making process under the complex scene of human drivers, a new method of driving behavior decision model is proposed. On this basis, the motion planning of unmanned vehicle is completed. The specific content of the research is as follows:
1) introduce the research significance of unmanned vehicle, understand the research status of unmanned vehicle at home and abroad, investigate the intelligent decision making and motion planning method of mobile robot, analyze and compare the realization way of the decision system of unmanned vehicle in foreign countries, describe and summarize the typical traffic situation of the city, and based on the driver's task. Demand and its decision-making process and behavior of vehicle, the key problem of decision planning of unmanned vehicle in urban road is put forward, and the design criterion of decision system is clarified. Then, each component module of "intelligent pioneer II" unmanned vehicle platform is introduced, and the working principle and cooperation mode of the platform are expounded. Based on the multi-resolution characteristics of time and space in the decision system, a framework of three layers of modular decision making system is designed to meet the real-time, adaptive and robust requirements of the decision system.
2) in view of the different driving environment and the driving behavior characteristics of human drivers, a hierarchical finite state machine is used to establish the behavior decision module of unmanned vehicle in the urban environment. The complex behavior of the driver is abstracted and decomposed, and the decomposition of the atomic behavior is used as the bottom state set of the state machine. In the decision-making process of complex drivers, a driving behavior decision model based on multi attribute decision-making method is proposed, which extracts the related attributes concerned by the driver during the driving process, judges and evaluates and obtains the final driving behavior pattern, making the behavior decision model conforms to the thinking process of the man type driver and solves the complex intersection of the city. Based on the entropy weight method, a driving behavior matrix weighting method based on entropy weight method is designed, which is based on the entropy weight method, and establishes a weight system based on the driving experience and objective data to reduce the interference of subjective randomness to the decision results, and to reduce the inaccuracy of the entropy method caused by the lack of sample data. In combination with two methods of TOPSIS optimization and grey relational analysis, a new grey ideal value approximation model of driving behavior is constructed to make decision evaluation, which makes the selected scheme not only close to the optimal scheme in space position, but also its shape is close to the optimal scheme, which ensures the optimal driving behavior.
3) the motion planning method based on radial basis function neural network is studied. First, the design principle of motion planning algorithm in urban environment is analyzed, and the difficulty of motion planning is clarified. A motion based on radial basis function (RBF) neural network is proposed for the characteristics of unstructured road features and unpredictable environment. In the planning method, the discrete reference points in the driving region are randomly extracted, the regularization network is used to approach the data, and the gradient descent method is used to study the network parameters with a single output RBF network learning method with a forgetting factor. The generated trajectory can fit the shape of any road and satisfy the vehicle transportation. In addition, because the RBF network is a local approximation network, it has the advantage of fast learning and fast response to the dynamic changing environment and meets the real-time requirements of the vehicle planning system.
Finally, based on the real road environment of the city, the experiment is carried out on the "intelligent pioneer II" unmanned vehicle platform. The results verify the correctness and effectiveness of the design method of the decision system.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：U463.6;U495
本文编号：2154888
[Abstract]:With the development of computer science and robotics, unmanned vehicles have been widely used in many fields, such as military, civil and scientific research. It focuses on the latest research achievements in many disciplines, such as structure, electronics, cybernetics and artificial intelligence, and has a broad application prospect.
Intelligent decision making is one of the key components of unmanned vehicle, which is one of the key parts of research. In urban environment, because of the complex and changeable driving scene, the behavior of the traffic participants is difficult to predict. It is impossible to use a standard and unified decision model to describe the unmanned vehicle in the complex environment of urban area. On the basis of learning the decision-making process under the complex scene of human drivers, a new method of driving behavior decision model is proposed. On this basis, the motion planning of unmanned vehicle is completed. The specific content of the research is as follows:
1) introduce the research significance of unmanned vehicle, understand the research status of unmanned vehicle at home and abroad, investigate the intelligent decision making and motion planning method of mobile robot, analyze and compare the realization way of the decision system of unmanned vehicle in foreign countries, describe and summarize the typical traffic situation of the city, and based on the driver's task. Demand and its decision-making process and behavior of vehicle, the key problem of decision planning of unmanned vehicle in urban road is put forward, and the design criterion of decision system is clarified. Then, each component module of "intelligent pioneer II" unmanned vehicle platform is introduced, and the working principle and cooperation mode of the platform are expounded. Based on the multi-resolution characteristics of time and space in the decision system, a framework of three layers of modular decision making system is designed to meet the real-time, adaptive and robust requirements of the decision system.
2) in view of the different driving environment and the driving behavior characteristics of human drivers, a hierarchical finite state machine is used to establish the behavior decision module of unmanned vehicle in the urban environment. The complex behavior of the driver is abstracted and decomposed, and the decomposition of the atomic behavior is used as the bottom state set of the state machine. In the decision-making process of complex drivers, a driving behavior decision model based on multi attribute decision-making method is proposed, which extracts the related attributes concerned by the driver during the driving process, judges and evaluates and obtains the final driving behavior pattern, making the behavior decision model conforms to the thinking process of the man type driver and solves the complex intersection of the city. Based on the entropy weight method, a driving behavior matrix weighting method based on entropy weight method is designed, which is based on the entropy weight method, and establishes a weight system based on the driving experience and objective data to reduce the interference of subjective randomness to the decision results, and to reduce the inaccuracy of the entropy method caused by the lack of sample data. In combination with two methods of TOPSIS optimization and grey relational analysis, a new grey ideal value approximation model of driving behavior is constructed to make decision evaluation, which makes the selected scheme not only close to the optimal scheme in space position, but also its shape is close to the optimal scheme, which ensures the optimal driving behavior.
3) the motion planning method based on radial basis function neural network is studied. First, the design principle of motion planning algorithm in urban environment is analyzed, and the difficulty of motion planning is clarified. A motion based on radial basis function (RBF) neural network is proposed for the characteristics of unstructured road features and unpredictable environment. In the planning method, the discrete reference points in the driving region are randomly extracted, the regularization network is used to approach the data, and the gradient descent method is used to study the network parameters with a single output RBF network learning method with a forgetting factor. The generated trajectory can fit the shape of any road and satisfy the vehicle transportation. In addition, because the RBF network is a local approximation network, it has the advantage of fast learning and fast response to the dynamic changing environment and meets the real-time requirements of the vehicle planning system.
Finally, based on the real road environment of the city, the experiment is carried out on the "intelligent pioneer II" unmanned vehicle platform. The results verify the correctness and effectiveness of the design method of the decision system.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：U463.6;U495
【参考文献】
中国期刊全文数据库 前10条
1 贺玉龙,卢仲贤,马国雄,刘小明,任福田;高速公路直线段车辆稳定运行速度模型[J];公路;2002年10期
2 郑海军;孙世群;何为;杨晨;;基于熵权的TOPSIS法在清洁生产和高费方案研选中的应用[J];环境科学与管理;2013年01期
3 任殿波;张继业;张京明;崔胜民;;智能车辆弯路换道轨迹规划与横摆率跟踪控制[J];中国科学:技术科学;2011年03期
4 高振海,管欣,李谦,郭孔辉;基于预瞄跟随理论的驾驶员跟随汽车目标速度的控制模型[J];吉林工业大学学报(工学版);2002年01期
5 张纯刚,席裕庚;动态未知环境中移动机器人的滚动路径规划[J];机器人;2002年01期
6 孙振平;贺汉根;;基于层次式的实时智能计算体系结构研究[J];计算机工程与科学;2006年04期
7 王刚;张成洪;黄丽华;;基于方法集的企业知识管理评价研究[J];科技进步与对策;2009年07期
8 孙树栋,林茂;基于遗传算法的多移动机器人协调路径规划[J];自动化学报;2000年05期
9 李灿;张凤荣;朱泰峰;奉婷;安萍莉;;基于熵权TOPSIS模型的土地利用绩效评价及关联分析[J];农业工程学报;2013年05期
10 燕明;李文清;周涛;;基于A*算法的坦克CGF全局路径规划[J];指挥控制与仿真;2008年03期
中国博士学位论文全文数据库 前3条
1 丛岩峰;高速公路环境中自主驾驶车辆运动规划与控制[D];吉林大学;2011年
2 孟庆梅;仿人头像机器人人工情感建模与实现的研究[D];哈尔滨工业大学;2009年
3 刘建国;虚拟人情感模型及认知方法研究[D];华中科技大学;2009年
，本文编号：2154888
本文链接：https://www.wllwen.com/kejilunwen/jiaotonggongchenglunwen/2154888.html
