【作者】 陈凤东；

【导师】 洪炳镕；

【作者基本信息】 哈尔滨工业大学 ， 计算机应用技术， 2009， 博士

【摘要】 近年来机器人加快了融入人类日常生活的步伐,这对机器人提出新的要求:不但具备安全自主导航能力,而且更加智能化和方便与人类交流。然而机器人目前的现状与人们的期望相去甚远。如何使得机器人具备更深层次的认知能力,是目前最大的挑战。本文深入研究了移动机器人环境感知、自主导航及目标跟踪等问题。首先在生物认知地图的启迪下,在移动机器人主流传感器-里程计、视觉和激光测距仪基础上,学习生物导航的优点及概率导航的成功之处,提出了基于多传感器联合顺序粒子滤波的认知地图创建算法,将移动机器人对环境的学习,从感知层次提高到认知层次;在认知地图基础上,提出一种基于认知地图的移动机器人自主导航方法:即在认知地图下实现了全局定位、位置识别、避障路径规划、基于位置的导航及精确末段泊位问题;最后,为增加移动机器人的环境交互能力,在认知地图下,提出三种有效的环境监测及目标跟踪方法,为机器人与环境交互提供技术基础。本课题来源于国家自然科学基金“基于不精确性地图的移动机器人室内导航技术研究(No.60643005)”、国家863基金“分布式多机器人合作与竞争机制及其应用技术(No.2006AA04Z259)”。主要包括以下内容:第一,研究了视觉尺度不变性特征(Scale Invariant Feature Transform, SIFT)的匹配方法;提出一种基于DD-BBF(Double-Direction Best-Bin-First)的SIFT特征匹配方法,有效解决了三维重建的数据匹配问题;研究了激光测距仪的极坐标扫描匹配(Polar Coordinates Scan Matching, PCSM)算法,增强了地图创建的可靠性。第二,研究了移动机器人基于多传感器联合顺序粒子滤波的认知地图创建算法:对传统RBPF(Rao-Blackwellized particle filter)和传统FastSLAM(Fast Simultaneous localization and mapping)进行改进,创建了视觉-激光测距仪联合度量地图。在度量地图创建过程中,把度量地图划分为许多子地图,子地图组定义为位置,并生成拓扑地图;利用导师知识,对位置进行语义关联(label),最终生成包含语义层-拓扑层-度量层的认知地图。第三,研究了基于认知地图的移动机器人导航方法:即在认知地图下,实现了全局定位、位置识别、避障路径规划、基于位置的导航及精确末段泊位。本文提出基于认知地图的导航方法,使得机器人获得了灵活鲁棒的大范围精确导航能力。对移动人机器人实际应用具有良好的指导作用,对解决移动机器人导航问题提供了一个新思路。第四,研究了基于认知地图的移动机器人目标监测和跟踪技术,以移动机器人认知地图作为基础,针对移动机器人本体感知范围有限的问题,采用视频感知网络扩展移动机器人的环境监测能力,提出一种背景差分算法,有效解决了照明变化条件下的目标检测跟踪问题,并通过单应矩阵实现了视频感知网监测信息与移动机器人认知地图的标定;研究了基于视觉特征的目标跟踪方法;对于人的定位和跟踪,提出一种基于视觉-激光测距仪的异质传感器顺序滤波算法,有效的实现了同时机器人定位和人跟踪,为机器人与环境交互提供技术基础。最后,介绍了基于认知地图导航的移动机器人硬件和软件系统。更多还原

【Abstract】 Robots are mending their pace to fuse into daily life of human being recently. Robots are expected to not only have ability of safe navigation but also are more intelligent and friendly to communicate with human being. But the state-of-the-art in mobile robotics is far behind expectation. The greatest challenge at the moment for mobile robotics is how to endow robots with the capacity to exhibit a greater degree of spatial awareness. The root-cause of the problem lies in the deficiency of semantic content in mobile robot representations. These problems form the core motivations of this thesis.The work presented here is an in-depth study of the problems of environment apperception, navigation and environment intercommunion. It studied the relative problems step by step. Inspired by biology cognitive mapping, a mobile robot cognitive mapping algorithm based on multi-sensor and joint-sequent particle filter is proposed by referring the strongpoint of biologic and probabilistic navigation and using the common sensors: odometer, vision and laser range finder. The algorithm can build a global, correlative, active map. Thirdly a navigation framework is presented based on the cognitive map: i.e. solving the problems of global localization, path planning and ending docking based on the cognitive map. Lastly three key technologies for intercommunion between the robot and the environment in the map are discussed which can build a reliable foundation for intercommunion between the robot and the environment.The problems studied in this paper derive from the National Natural Science Foundation of China“Research on Mobile Robot Indoor Navigation Technology Based an Inaccurate Map”, and the National 863 High-tech Research and Development Plan of China“Collaboration and Competition Mechanism for Distributed Multi-robots and its Application Techniques”. The contributions are discussed in detail as follows:Firstly, the SIFT(Scale Invariant Feature Transform) algorithm which is used to represent environment is discussed, and a DD-BBF(Double-Direction Best-Bin-First) based SIFT feature matching method is presented to solve the data matching problem in 3D reconstruction of space efficiently; A PCSM(Polar Coordinates Scan Matching) method is discussed for the robust laser data matching.Secondly, a mobile robot cognitive mapping algorithm based on multi-sensor and joint-sequent particle filter is studied.As to build the cognitive map, an improved RBPF(Rao-Blackwellized particle filter) and FastSLAM are used to implement a jointSLAM. A more precise vision-laser joint metric map is built by this method. When building the joint metric map, the map is partition into many groups of submaps by a special partition rule. Each submap group is defined as a place, and every place is encoded with semantic conception which comes from a tutor semantic sequence. When the semantic map is built, the cognitive map is founded at the same time.Thirdly, a navigation method based on cognitive map is studied to solve the problem of global localization, HMM based place recognition, path planning, place based navigation and precise docking. The highlight point of the method is that it can well instruct the real robot application, and endow mobile robots with the abilities of flexible and precise navigation in large scale environment. It also provide a new thought of solving the navigation problem.Finally, object surveillance and tracking method based on the cognitive map are discussed, and a vision network with new background subtraction for non-stationary scenes is used to extend the observation range of the mobile robot. A homography matrix is used to settle the calibration relationship between the robot and the vision network; a monocular vision based object following method is presented which uses SIFT to judge the distance and direction of the object; as to human localization and tracking, a heterogeneous sensor sequence filter based on laser-vision is presented to locate and track the people efficiently. These works can build a reliable foundation for intercommunion between the robot and the environment. lastly, the hardware and software of the system are introduced.更多还原