【作者】 郑建冬；

【导师】 张丽艳；

【作者基本信息】 南京航空航天大学 ， 航空宇航制造工程， 2013， 博士

【摘要】 三维坐标数据测量是逆向工程、机器人视觉自主导航、产品检测、虚拟现实等的首要步骤。本文以提高基于单数码相机的三维摄影测量的稳健性和精确性为研究目标，对测量目标的图像提取、多位姿相机自定位、三维重建中外点剔除、三维坐标数据全局最优解、系统参数辨识与校正等关键问题进行了深入研究。本文的主要研究内容和成果概况如下：1.提出了基于Canny边缘检测算子和多项式曲面拟合的亚像素图像边缘提取方法，为实现圆形编码点中心的精确定位提供了良好的图像数据基础。编码点中心的图像坐标定位精度是影响三维测量系统精度的关键因素之一，在亚像素边缘提取基础上，提出一种改进的灰度重心方法计算编码点中心，该方法能实现编码点中心的快速、精确图像坐标定位。与传统的最小二乘椭圆拟合法、灰度重心法以及近年来提出的基于对偶二次曲线拟合方法进行比较分析，验证了本文方法具有更高的稳定性和定位精度。2.提出了一种改进的确定两幅图像相对拍摄位姿的鲁棒估计算法，解决了本质矩阵多解问题，修正了鲁棒估计算法中的参数估计公式。与当前具有代表性的两种5点算法进行定量比较研究，实验证明本文算法的鲁棒性得到提高，为后续整个图像组相机位姿及编码点三维重建的稳定性奠定了很好的基础。结合相机全局旋转运动估计理论，提出了基于RANSAC方法的多视图相机全局坐标旋转运动估计方法，求出各相机在统一世界坐标系的全局旋转运动。运用图论中最大生成树方法构造两幅图像之间编码点匹配准确性最高的图像组，以建立编码点的全局匹配关系。3.提出了一种快速、稳健的已知相机全局旋转姿态的多视图几何重建方法。该方法通过对空间点再投影误差函数进行线性化分析，提出了在三维重建过程中基于L1范数最小化和基于最小最大化问题的自适应外点剔除方法，通过引入松弛变量，将外点剔除问题转换为关于松弛变量的线性规划模型，一方面提高了算法的稳定性，另一方面提高了算法的效率。与三种具有代表性的基于松弛变量的外点剔除算法进行比较实验，结果表明该方法不会出现过度剔除外点的现象，能正确保留真正的内点，并且具有更小的空间点再投影误差。4.根据凸优化理论，给出了在L2范数下多视图三维重建结果的全局最优判定方法，该方法通过判定空间点再投影误差函数的凸性，来确定已经求出的最优解是否为全局最优解。对于非凸情况，运用分支定界理论，提出了求解非凸函数比式和问题的空间点三维坐标全局最优解的方法。通过与基于L￥范数全局优化方法的比较实验证明，本文方法具有更高的计算精度和运算效率。5.提出了一种融合简明隐式图像校正的多视图几何精确重建方法。该方法引入6个隐式畸变参数进行镜头校正，并将隐式畸变参数直接嵌入到三维重建优化模型中。实验表明这个算法能够显著提高三维摄影测量精度，在图像组多视图几何精确重建中起着关键作用。6.在精确的多视图几何求解架构基础上，提出一般特征点和特征线的三维摄影测量方法。首先研究了两幅图像之间的特征点匹配算法，该算法在编码点的全局匹配关系下，采用形状上下文描述算子和概率松弛标记法实现了特征点的一对一匹配，再运用薄板样条函数对匹配点集的整体形状进行匹配验证。然后提出并实现了特征线提取、特征线全局匹配及特征线三维重建的方法。完成了汽车车门钣金零件、飞机机翼、管路模型架、摩托车骨架等实物的三维测量。本文提出的算法和系统软件模块均进行了实验验证、对比和分析，应用原型系统测量了具有代表性的三维实物，获得了良好的测量稳健性、较高的测量精度和测量效率。更多还原

【Abstract】 Three-dimensional (3D) coordinate measurement is prerequisite for reverse engineering,product quality inspection, robot visual navigation, virtual reality etc. The work in this thesisfocuses on improving the robustness and accuracy of digital close range photogrammetrymeasurement system based on a hand-held CCD camera. Several key issues in3D measurementsystem are further studied in this paper, including target extraction from image, multi-cameraposition and orientation locating, outlier removal in3D reconstruction, globally optimal solutionof3D coordinate, system parameter identification and correction etc. The main contents andcontributions of the dissertation are as follows:1. An edge extraction method based on Canny edge detector and polynomial surface fittingwith sub-pixel accuracy is presented. It provides a good foundation for precise location of thecircular coded marker centers. The precision of the image coordinates of the coded markercenters is one of the critical factors for3D photogrammetry. An improved grey centroid methodis put forward in the paper. Compared with the least-squares ellipse fitting, the traditional greycentroid and the dual conic fitting approach presented recently, the proposed method can obtainsignificantly more stable and precise coded marker centers.2. An improved algorithm for robust estimating the camera relative position and orientationof pairs of views is put forward. The multi-solutions problem involved in the essential matrixestimation is resolved. In addition, the parameter estimation formula in the robust model iscorrected. Comparisons with several representative five-point algorithms show that the algorithmcan advance the robustness and precision of the camera relative orientation. The algorithm serveswell as the input for the following3D reconstruction pipeline. By utilizing the rotationalconsistency theory, an approach to estimating the global rotations in multi-view reconstruction isproposed. It achieves the camera relative orientations under the world coordinate system.According to the maximum spanning tree algorithm in graph theory, image group with thehighest reliability of coded markers correspondences between image pairs is built. A globalmatching relationship can be achieved by the reliable image group.3. A fast and stable method for multiple view geometry reconstruction with known rotationsis presented. Using the linearity of re-projection error function, this article puts forward theadaptive outlier removal method usingL1norm and minmax problem during3D reconstruction.By introducing slack variables into the objective function, the outlier removal problem can beanswered by the linear programming problem about slack variables. It improves the stability and efficiency of the algorithm simultaneously. Comparisons with three representative outlierremoval approaches based on slack variables show that the method in this paper can keepgenuine inliers and decrease the re-projection error.4. According to the convex optimization theory, an approach to verifying global solution formultiple view geometry reconstruction problems usingL2is derived. The convexity of there-projection error function is utilized to verify whether the local solution is globally optimal. Abranch and bound algorithm for globally solving the non-convexity sum of ratios problem is putforward. Compared to L￥-norm-based global optimization, the proposed approach has highercalculation accuracy and computing efficiency.5. A concise implicit image correction approach for accurate reconstruction of multiple viewgeometry is presented. By introducing six implicit distortion coefficients, the image correctionapproach has been integrated into the3D reconstruction optimization model. Experiments haveevidenced that the method can significantly improve the3D positioning accuracy. The methodplays a key role in accurate multiple view geometry reconstruction.6. Based on the established accurate multiple view geometry, the methods for3Dreconstruction of points or line segments are put forward. Point correspondences between imagepairs are generated. Under the global correspondence of the coded markers, the match for twopoint set is obtained using shape context and probabilistic relaxation. Then the matching result isverified by thin plate spline function. The method for line segments feature extraction, globalcorrespondence, and3D reconstruction is proposed. Practical examples include an automobiledoor sheet metal part, the wing of an aircraft, a pipeline model frame, motorcycles and othersample objects.The proposed algorithms and system modules have all been fully experimented, comparedand analyzed. It has been demonstrated that the representative target objects can be obtained bythe prototype system with good measurement stability, high measurement accuracy, andsatisfactory measurement efficiency.更多还原