【作者】 杨淼；

【导师】 魏志强；

【作者基本信息】 中国海洋大学 ， 地图学与地理信息系统， 2009， 博士

【摘要】 计算机视觉既是工程领域,也是科学领域中的一个富有挑战性的重要研究课题。基于计算机视觉的三维信息获取技术已成为当今的研究热点。本论文主要研究基于图像的三维自动重建的参数化建模方法,试图从数码相机拍摄的建筑物图像中以参数化建模的方法恢复建筑物的三维几何结构,从而为各种应用环境提供有效的建模方法和三维模型,便于浏览,可移植性强。在分析了目前国内外已有的研究成果基础上,进行了多方面的深入研究和实验,取得了一些有意义的成果。图像分割是计算机视觉应用中非常重要的预处理步骤。本文在总结彩色图像分割算法的基础上,根据建筑物彩色图像分割的要求和特点,针对分水岭变换存在的过分割问题,提出了基于分水岭变换和区域融合的彩色建筑物图像分割算法,该方法结合区域面积控制预处理,利用建筑物的轴对称特点设计融合代价函数,并设计了根据代价函数值的直方图分布,自动确定区域合并停止阈值的方法。另外,还提出了一种基于尺度空间图像分析的彩色图像分割算法,该算法采用形态学分层图像分析的方法,由体积灭值构建图像尺度空间的树状表达。通过对尺度空间的生存时间属性和其它特征的分析,实现建筑物图像分割。形态学尺度空间(形态筛)是一种近年来广泛应用于图像处理、分析中的动态分析框架,可以更容易获得图像的本质特征。在图像分割预处理之后,本文着重讨论了形态筛在彩色图像中的扩展应用,在对比分析两种现有彩色形态筛的基础上,提出了一种新的基于模糊模型的彩色形态筛,并验证了模糊彩色形态筛各方面的性能。另外,本文将灰度形态筛引入到基于图像的建筑物参数化建模的各个过程中,发挥不同尺度形态筛的作用,取得了较好的效果。本文为城市建筑物建立了面向对象参数化数据模型,提出了面向对象的CSG点、面、体分层建筑物建模方法,并根据模型的特征研究了自动实现建筑物类型判断的算法,该方法首先应用形态筛保留建筑物的主要轮廓区域,然后从图像的点、线特征入手,利用Hough变换提取轮廓线,并提出了一种线段优化算法,这种优化算法可有效排除Hough变换和线段整合过程中产生的错误线段,并用于剔除错误的轮廓线交点和建立角点邻接图。最后,根据提取的各类角点的数目进行建筑物类型的判断。算法可自动对单幅图像中的规则建筑物进行平顶或非平顶的判断。城市建筑物的主体以平行六面体居多,而获得建筑物的轮廓是实现单幅图像重建的关键,本文设计了一种城市建筑物贝叶斯概率统计模型来模拟城市建筑物世界,在假设相机内参已知的情况下,利用边缘像素通过最大化后验概率(MAP)估计消隐点的投影,而后用大尺度形态筛结合估计到的消隐点对边界像素进行分类并自动提取建筑物主体轮廓。在全面而深入地研究平行六面体的几何特征与相机参数之间的关系后,根据获得的角点实现平顶建筑物的三维参数自动提取。现有的大部分单幅图像的重建方法都不是根本的参数化建模,也没有对建筑物类型判断的过程。本文针对未知场景的情况,研究了仅由交互输入的六个角点,实现非平顶建筑物屋顶角点的自动提取,进行两种常见非平顶建筑物模型的判断,并实现三维参数化建模的方法,参数化重建省去了平面纠正的步骤,不必重建大量的点或直线,也不需要用户指定平行性、共面性等约束,且同时完成相机的标定和模型参数的确定,与那些多视图的自动重建方法相比,更加可靠和简单。论文中将提出的各种方法和算法分别应用于模拟和真实图像进行实验。最后实现了Web页面中连同VRML技术的建筑物场景的三维建模和漫游及建筑物相关属性的显示。更多还原

【Abstract】 With the development of information science, computer vision technology has been widely applied to many areas. The issue of 3-D information acquisition techniques based on computer vision has been one of the research hotspots in this field. The main idea of this thesis contains monocular vision techniques for estimating geometric properties of 3-D world from a series of 2-D digital images. Based on previous researches, significant progresses have been achieved after deeply study and analysis of former methods.Image segmentation is a necessary pretreatment step in many computer vision applications. In this thesis, a segmentation method especially for color images contain buildings is proposed. This approach makes use of area control pretreatment and the watershed algorithm to produce the original regions. The axial symmetry combined with other properties is estimated as region dissimilarity features. The final segmentation is derived by merging process. And the termination criterion determined through the distribution of the merging costs is an adaptive threshold. In addition, another segmentation algorithm based on scale-space image analysis is proposed. It analyzes hierarchical, geometric properties and the hue uniformity measure of the tree of scale-space in order to segment the objects. The results show that the approaches proposed can obtain effective regions of buildings in original images.The morphological sieve is a pervasive dynamic analysis frame since it can easily obtain the substantive characteristics of images. In this thesis, a new color morphological sieve with better synthetic performance proved by the experiments based on fuzzy theory is proposed. Besides, the morphological sieve is introduced into the monocular vision reconstruction process which results in a sufficient use of characteristics of the different scale sieves and good modeling results. After constructing the object-oriented parametric data model for the city building, the corners’properties are analyzed and the automatic building model estimation is realized. In the whole process, an effective segment optimization method is combined to improve the accuracy of the line and corner extraction.Automatic modeling of 3D object from images is a challenging problem in areas of computer vision especially from a single image. In this paper, a“City Building World”Bayesian model is constructed and a new automatically parameterized reconstruction method is presented which can recover the 3D cuboid architecture structure from a single terrestrial image. With the assumption that the interior parameters of the camera are known, the projection of the vanishing points is determined by exploiting the MAP. And from a few boundary pixels, the main parallelepiped contour is estimated. The cuboid model is reconstructed from the relationship between the parallelepiped geometry and the camera parameters.A parametric modeling method for the non-flat roof building is also presented. By using only six interactive input points, the roof corners can be detected automatically as well as the model geometry parameters. The proposed approaches and algorithms in this thesis are applied to synthetic data and real images respectively. At last, the rendering and virtual roaming dynamic web page is designed.更多还原