【作者】 李海丰；

【导师】 刘景泰；

【作者基本信息】 南开大学 ， 控制理论与控制工程， 2012， 博士

【摘要】 自主定位技术是移动机器人自主导航的基础和关键。近年来，随着图像传感器生产成本的降低，以及图像处理、模式识别等相关技术的迅速发展，视觉定位受到越来越多的关注。移动机器人视觉定位是指借助于摄像机采集的视觉信息，确定机器人的当前位置。近年来，移动机器人在户外环境下，特别是城市区域中的应用越来越广泛。因此，城市环境中的移动机器人视觉定位研究具有非常重要的理论意义和极其广泛的应用前景。论文围绕城市环境中的移动机器人视觉定位问题展开研究，实现了城市环境中无准确GPS数据情况下的移动机器人精确定位。论文提出了利用机器人工作区域的卫星地图和移动机器人车载相机所采集的图像二者相结合的视觉定位策略。其中，卫星地图用来提取周围建筑物的俯视轮廓，而摄像机图像用于重建图像中的建筑物轮廓，通过对二者进行匹配来确定移动机器人在二维卫星地图中的绝对位置。为了实现上述定位过程，论文设计了基于线特征的场景三维重建方法。传统的进行场景三维重建的方法主要是利用点特征。基于点特征的三维重建方法存在着精度低、计算量大以及无法准确地表示当前场景等问题。与点特征相比，线特征存在诸多优点：在同样的噪声强度下，线特征受噪声影响更小；线特征对光照情况和阴影都不敏感；线特征的数量较少，从中确定有用线段并利用其进行三维重建的计算量较小。然而，不同视图间的线特征匹配一直都是计算机视觉领域的一大难点，至今仍缺乏能够准确、全面地查找线对应的方法。为此，论文首次提出了多层特征图（Multilayer Feature Graph, MFG）结构。MFG借助于多种特征之间的几何关系和约束，可以有效地确定不同视图间线特征的对应关系，并最终确定线特征和建筑物平面（竖直平面）的三维信息。同时，MFG还是一种有效的场景表达的方式，将周围场景表示为多种相互关联的主要特征，包括点特征、线段特征、直线特征和竖直平面特征，更有助于对场景的理解。论文的主要研究内容包括MFG的设计与构建、基于MFG的视觉定位算法两部分。MFG的设计与构建部分主要包括MFG的结构设计以及基于特征融合的MFG的构建方法。视觉定位算法部分主要讨论了如何利用高分辨率的卫星地图以及MFG确定移动机器人在卫星地图中的准确位置。首先设计并实现了利用卫星地图自动提取俯视的建筑物二维轮廓的方法。然后提出了利用单个MFG和建筑物俯视轮廓的特征加权视觉定位方法，实现了移动机器人的简单、快速定位。但该方法的缺点是无法保证定位解的唯一性和正确性，特别是在机器人所处的环境中存在很多相似建筑的情况下。为此，论文又设计了基于投票的视觉定位方法。基于投票的定位方法同时利用了多个MFG，每个MFG提供若干个候选解，最后根据候选解的一致性来确定最终解。论文的整体工作概述如下：(1) MFG的设计与构建。给出了MFG的模型结构以及特征的提取方法。利用MFG中多层特征之间的几何关系，提出了基于特征融合的MFG的构建方法。通过构建MFG实现两视图间线对应的查找，并实现了基于线对应的场景重建与理解。(2)基于高分辨率卫星地图的建筑物轮廓自动提取。针对高分辨率卫星地图中建筑物以及非建筑物区域的特点，同时借助相应的普通的城市电子地图，提出了一种自动、快速的建筑物二维轮廓信息提取方法。利用Google卫星地图对该方法进行了实验验证。实验结果表明，该方法可以快速、准确、自动地完成建筑物轮廓提取。(3)基于单个MFG的特征加权视觉定位。利用单个MFG以及二维的建筑物俯视轮廓信息，提出了基于特征加权的视觉定位方法。该方法将定位问题转化为一个优化问题，通过求解该优化问题来定位机器人。实验表明，该方法在大多情况下能够实现移动机器人的快速自主定位。然而，理论分析和物理实验均表明，当机器人所处的环境较复杂，特别是周围存在很多相似的建筑物时，该方法无法保证定位解的唯一性，甚至可能出现错误定位。(4)基于投票的视觉定位。该方法是对基于单个MFG的特征加权视觉定位方法的改进。基于投票的视觉定位方法利用在相同位置采集的多组两视图构建的多个MFG，根据特征加权的方法，每个MFG可以提供若干个候选解，最终由所有的候选解进行投票来确定最终解。实验结果表明，该方法可以有效地降低错误定位的概率，同时可以提高定位的精度。更多还原

【Abstract】 Self-localization is a basic and key technique for mobile robot navigation. Inrecent years, with the reduction of camera cost and the rapid development of thetechniques for image processing and pattern recognition, researchers pay moreattention to the visual localizaton (also termed as “vision-based localization”). Visuallocalization for mobile robot utilizes the visual information from the camera(s) tolocalize the mobile robot. Recently, the applications for mobile robots working inoutdoor environments, especially in urban areas, are very popular. Therefore, theresearch on visual localization for mobile robot in urban environments is veryimportant both in theoretical and practical applications.The dissertation focuses on visual localization for mobile robot in urbanenvironments. A localization scheme combining the satellite map of mobile robotworking area and images captured by an onboard camera is proposed. The satellitemap is used to generate the building boundaries from top-down view, and the cameraimages are utilized for building reconstruction from horizontal view. The proposedmethod can determine the absolute position of the robot in the2D satellite map bymatching the two results obtained in the above.To realize the localization scheme above, a line-based3D reconstruction methodis proposed. The point-based3D reconstruction methods are popular. However,existing point-based methods are low-accuracy, high-computation, and can notrepresent the scene exactly. Compared with point features, line features are morerobust, and insensitive to lighting condition or shadows. The line-based methods canlead to low computational cost because of their small amount and robustness.However, the matching of line features between views is a difficult issue in computervision. To this end, a structure named Multilayer Feature Graph (MFG) is firstlyproposed in this dissertation. With the aid of the geometric relationships betweendifferent features, MFG can find the line correspondences between two viewssuccessfully, and reconstruct line features and vertical planes in further. Besides, MFG is an effective method to facilitate the robot scene understanding byrepresenting the scene as different related key features, such as points, line segments,lines and vertical planes.The main work consists of two parts: the design and construction of MFG,localization algorithms based on MFG. A feature fusion method is discussed in theMFG construction part. While in the localization algorithms, an automatic buildingboundary generation method from high-resolution satellite map is developed firstly,then a feature weighted localization method based on an MFG is proposed. However,this method can not guarantee the correctness and uniqueness of solution. Thus, avoting-based localization algorithm based on multiple MFGs is designed, where eachMFG could provide several candidate solutions, and the final solution is determinedbased on the consensus of all the candidate ones. In general, the main work in thedissertation is summarized as follows:(1) Design and construction of MFG. The structure of MFG and the extractionmethods for the features in MFG are proposed. A feature fusion method for MFGconstruction is developed based on the geometric relationships between features inMFG. MFG can help us to find the line correspondences between two views, usingthat we can realize the scene reconstruction and understanding well.(2) An automatic building extraction method from high resolution satellite map.By analyzing the characters of building and non-building regions in satellite map, anovel automatic and time-efficient building boundary extraction method, with the aidof corresponding ordinary map, is proposed. This method was implemented andtested on the Google satellite maps. The physical experiments demonstrated theaccuracy and efficiency of the method.(3) Feature-weighted map query method for localization. A feature-weightedvisual localization method is proposed based on a single MFG and2D buildingboundary map. This method converts the localization task into an optimizationproblem. The location solution can be obtained by solving the optimization problem.The physical experiments demonstrated that the method can localize the robotsuccessfully in most cases. However, the theoretical analysis and physicalexperiments show that, in some complex situations, especially when there are similar buildings in the surroundings, this method can not guarantee the uniqueness of thesolution, even leads to wrong results.(4) Voting-based visual localization method. This method is an improvement ofthe feature-weighted algorithm above. Multiple MFGs generated from camera framesare utilized in the voting-based method, with each MFG providing multiple candidatesolutions. The final localization solution is determined based on the consensus of thecandidate ones. Physical experiments demonstrated that, compared with the featureweighted method, the voting-based algorithm can improve the probability ofcorrectness and the localization accuracy.更多还原