计算机视觉三维重建理论与应用

【作者】 王宇宙；

【导师】 赵宗涛；

【作者基本信息】 西北大学 ， 计算机软件与理论， 2004， 博士

【摘要】 计算机立体视觉研究利用二维投影图像恢复三维景物世界的问题，是计算机视觉技术的主要任务之一。计算机三维重建理论及其实现技术在战场三维地形建模、战场监视以及精确导航等领域有着重要的应用。根据军队建设对快速三维地形重建和快速精确定位的需要，本文对立体视觉计算中的影像特征提取、基于知识的影像分析、宽基线影像匹配、影像地面控制点自动识别、三维实时重建以及立体视觉模型误差分析等问题展开了理论和实现技术的研究，取得了一些研究成果。 图像基本特征获取是计算机三维重建多种算法的基础。通过对直线特征的分析，本文提出了一个新的层次结构化直线提取算法。该算法在不同尺度上利用局部区域直线特征构造直线段，形成一种具有层次的结构化的直线表示。算法采用自底向上的策略，首先构造底层局部线段结构，然后利用业已获得的下层结构构造上一层结构，因此具有较高的效率。算法既利用了局部特征，又利用了全局特征，因此具有较高的可靠性。 在区域特征提取方面，本文的贡献是提出了一个稳健的区域分割算法。该算法首先将影像划分为较为平滑的区域和亮度变化较为剧烈的区域。根据像点是否处于平滑区域采用不同的方法分别处理。对于处于平滑区域的像点采用简单算法处理，对于可能处于区域边缘的像点采用较为复杂的移动平均算法在色度空间和尺度空间上处理。因此，算法兼顾了时间效率和稳健性。 影像匹配是三维重建问题的核心。本文提出了一个面向三维重建任务的基于知识的图像分析系统模型。该模型将相关知识划分为基本影像处理知识、与具体目标影像相关的知识和与分析任务相关的知识。基于知识的图像分析系统运用基本影像分析知识和领域专家知识对场景中基本特征关系进行分析，获得场景中透视关系或其它特定空间几何关系的定性表达。目的是通过正确解释影像中某些空间关系来提高匹配算法的效率和可靠性。 在影像匹配算法方面，本文的主要贡献是提出了一个基于局部仿射不变量特征的宽基线条件下的影像匹配算法，处理最一般情况下的影像立体匹配。算法以影像特征点为定位点，使用分层变尺度窗口内的几何仿射不变量特征和亮度仿射不变量特征实现宽基线影像立体匹配。该算法利用局部特征进行宽基线立体匹配，因此具有较高的算法效率。同时，算法在较大窗口范围内构造尺度、旋转不变特征，并考虑到可能存在的投影变形而对窗口形状进行变换，因而具有较高的匹配可靠性。通过将大窗口划分为较小的重叠子区域，算法可以获得较好的匹配精度并减少因遮挡造成的不利影响。 在三维重建方面，本文研究了三维射影重建和欧氏重建的基本算法、光束法平差三维重建算法和实时Kalman滤波修正三维欧氏重建算法。探讨了基于平面图的单视图约束和多视图约束三维重建算法。 在立体视觉模型方面，本文推导了标准立体视觉模型和汇聚立体视觉模型的三维重建误差公式，分析了欧氏重建中不同立体模型的点位误差大小及其变化规律。 在计算机三维视觉理论与技术的应用方面，本文提出了一个快速数字地形重建系统的设计与实现方法。该系统将交互式系统和实时系统集成在一起，利用交互式三维重建系统获取高精度、高可靠性的三维模型，利用实时三维重建系统获取精度较低的快速重建结果，通过数据融合消除不可靠的重建结果、提高重建精度。该系统由控制测量子系统、交互式三维地形重建子系统、三维地形可视化子系统、基于GPS八NS的实时三维地形重建子系统和基于平面图的三维重建子系统等多个部分构成，实现高精度快速三维地形重建任务。该应用系统的设计和部分实现，进一步证明了本文理论研究和应用研究成果的正确性和实用价值。更多还原

【Abstract】 Computer stereo vision mainly studies the issues of reconstruction of the 3D world from the 2D images, and is one of the main tasks of computer vision. Computer stereo vision theories and techniques can be used in many important areas such as 3D battlefields modeling, battlefields situation monitoring, positioning and navigation. To meet the requirement of the military development for the 3D terrain reconstruction and positioning rapidly and precisely, this dissertation mainly studies the issues of image features detection and knowledge based image features spatial relation analysis, wide baseline stereo matching, real-time 3D reconstruction and error analysis about the stereo vision models, and proposes some novel and effective methods to cope with those problems.Finding out basic features in an image is always the necessary step in many algorithms in computer stereo vision. After studying the property of the straight line, a novel hierarchical algorithm about straight lines extraction is proposed in the dissertation. This straight lines extraction method extracts local line segments from local image intensity features at different spatial scale. First, the algorithm forms many short lines in the lowest level. Then it merges them into upper level according to their adjacency and straight line constraints. Finally a hierarchical straight lines structure is formed and final straight lines were extracted.In the area of region features detection, another contribution in this work is that it proposes a robust segmentation algorithm. This algorithm uses different methods dealing with different pixels in the image according to its position, namely using simple average method dealing with the pixels which are in the position of distance to edges of regions, and using the mean shift algorithm dealing with the other pixels, which are possibly near the edges of regions, both in color space and scale space. It is obvious that the proposed algorithm is effective and robust.The key problem to 3D reconstruction is image correspondence. A knowledge based image analysis system framework is proposed in the work aiming to deal with the imagematching problems in 3D reconstruction. In this framework, knowledge can be classified as knowledge about image processing itself, knowledge about specific target images and knowledge about image processing tasks. Knowledge based image analysis system employs fundamental knowledge about image analysis and experts’ knowledge in specialist area to find out spatial perspective relations among interested features in the scene or other relations such as perspective vanished point position and 3D parallel lines.In this work, a novel algorithm based on local affine invariants is proposed for searching correspondences between images in the most general case, namely, under the wide baseline conditions and without any priori information about the internal or external camera parameters. This algorithm uses image point features as anchor points for image matching, and the local space-variants regions around the anchor point are used as local window region to generate geometric affine invariants and color affine invariants for image matching. The proposed stereo matching method is efficient in algorithmic time because only the local image features are used in the image matching process. Meanwhile, the image matching results are reliable and robust because its feature matching region is relatively large, and the shape of local window region is transformed according to possible projective deformation. In addition, overlapped sub-regions are used in local feature extraction to reduce partially occluding problems.On 3D reconstruction, this dissertation discusses the basic methods of 3D projective reconstruction and 3D Euclidian reconstruction, a bundle adjustment algorithm for 3D reconstruction, a real-time Kalman filtering 3D Euclidian reconstruction method, and plane based single view and multi-view constraints algorithm for 3D Euclidian reconstruction.In this work, stereo camera model error is al更多还原