【作者】 陈泽志；

【导师】 吴成柯；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 2002， 博士

【摘要】 本论文研究了如何由非定标图像序列重建三维实体模型，并对由非定标图像序列测量三维物体中的若干关键技术进行了深入研究。传统的方法都是针对已定标的图像序列进行的，因而很难推广应用，实用性较差。本文的重点是在理论和实践两方面研究了，在既不知道相机内部参数，又不知道相机外部参数的情况下，如何由图像序列进行三维欧氏重建并最终获得三维欧氏测量信息的问题。本文所给的算法增强了三维实体模型的真实感，提高了重建三维模型的速度，测量精度高。主要成果如下： 1．给出了双三次B—样条平滑滤波算子，并且利用该算子进行了大量的边缘检测实验。实验结果表明，所给的方法对边缘检测非常有效，具有良好的鲁棒性。为图像的稠密匹配奠定了坚实的基础。 2．给出了两种高精度估计基础矩阵的线性算法： 1)．加权归一化算法。首先以8点算法的结果作为初始值，计算每对匹配点的余差和权因子，并将原始输入数据加权归一化处理，然后再用8点算法求F阵的8个参数，实现了F阵的估计。 2)．加权平移算法。首先将原始输入数据加权，计算加权后数据的重心坐标，并将坐标原点平移到该重心坐标，再作归一化处理，然后用8点算法求出F阵的8个参。大量的模拟数据和真实图像的实验结果表明，这两种算法不仅计算速度快，具有良好的鲁棒性，而且还提高了基础矩阵的估计精度，其鲁棒性和精度都明显优于广泛使用的改进的8点算法。 3．提出了新的基础矩阵估计非线性算法。首先用无约束规划求出双对极点位置，其目标函数是一个4元6次多项式，然后用SVD方法求出其余4个参数，实现了F阵的稳定估计，有效地避免了传统的用非线性最小二乘法求解所存在的不足。实验结果表明，所提出的方法彻底克服了原双对极点约束方法运算速度慢、存在大量伪解和对极点的不稳定现象，用较少的匹配点即可获得较高精度的F矩阵，实用性较强，且具有明显的几何意义。 4．提出了一种新的基于线性模型的摄像机自定标方法。利用三点透视投影图、灭点和向量正交的性质，得到一组与内参有关的非线性方程组，并将其转换为线性方程组，避免了求解过程中的累积误差，最后高精度地求出了全部内参α_x，α_y，u₀，v₀。并以此做初值，综合景物的先验知识，如正交直线（平面）、平行直线（平面）等求解绝对二次曲面，减少了求解绝对二次曲面的二义性，使绝对二次曲面的仿射特性和欧氏特性更精确，从而使重建的三维景物更为精确。 5．提出了新的图像矫正算法，使稠密匹配沿水平线进行，不但能简化匹配过程，而且还能提高匹配精度。此方法避免了相机定标而直接采用Bresenham算法沿对应对极由非定标图像序列重建和测量三维物体 线提取象素，将重构图像时象素丢失的可能性降到了最低限度，执行速度快，重构 图像小，且能适应相机的各种位置关系，大大提高了图像稠密匹配的精度，为景物 的高精度三维重建奠定了基础。6.提出了一种新的基于手提相机的由二维图像序列进行三维欧氏重建的非接触式测量 方法。与传统的测量方法相比，所给的方法具有对于摄像机不需要任何关于内参和 外参的先验知识、使用的设备简便通用、测量精度高以及技术易于推广应用等特点。更多还原

【Abstract】 This thesis discusses the problems of the recovery of a realistic textured model and the critical issues in metrology of 3D objects from uncalibrated image sequences. Traditional approaches are based on a preliminary calibration of the camera setup. This, however, is not always possible or practical. The goal of this work is to investigate the theoretical and practical feasibility in metrology of 3D objects from image sequences without any prior knowledge either about the parameters of the cameras, or about their motions. Meanwhile, several algorithms, such as how to improve the speed and the accuracy of reconstructed 3D model, are considered in this thesis. Main contributions are as follows:1. The bicubic B-spline smoothing filter operator is introduced. At the same time, a method for detecting the edge by using this operator is presented. The experimental results show that the method is efficient in edge detection, and it is very robust to noises.2. Two linear algorithms for estimating fundamental matrix are presented.i) Weighted normalizing algorithm. Weighted normalizing original input matching points with a weight factor related to residual errors and calculated the eight parameters of fundamental matrix by using the 8-point algorithm. The experimental results show that the algorithm is very robust to noises and outliers, and the fundamental matrix with high accuracy can be found.ii) Weighted translation algorithm. Normalized original input matching points and calculated the centroid coordinates by exploiting the strategy of weighted translation transformation, and the origins of coordinates are translated to their centroids. The eight parameters of fundamental matrix can be solved by using the 8-point algorithm and the procedure of estimating F-matrix with high accuracy can be achieved.3. A new method, the biepipole constraint algorithm, is developed to estimate the fundamental matrix (F-matrix) based on an 8-parameter model and the geometrical analysis. First, through the analysis of the new constraints, the four parameters of the F-matrix can be estimated by solving a nonlinear unconstraint optimization problem. The objective function of the optimization problem is an equation of degree six in four unknowns. Then, the four other parameters of the F-matrix can be evaluated by using the SVD method. Particular novelties of the algorithm are the obvious geometrical meanings of the parameters, fewer matching point pairs and higher accuracy.4. A new self-calibration method is presented. A set of equations is got by using the characteristic of 3-point perspective projective, vanish points and the orthogonal vector and then all the intrinsic parameters ax,ay,u0,vQ can be solved with high accuracy. This method also provides a credible initial value for nonlinear self-calibration method.In self-calibration, the prior knowledge of orthogonal/parallel lines and orthogonal walls are formulated as constraints on absolute quadric. The ambiguity on absolute quadric is reduced much more. Thus, the projective reconstruction and Euclidean reconstruction are more accurate and more realistic.5. A novel and efficient image rectification method using the fundamental matrix is proposed. In this approach, camera calibration is not required, and image resampling becomes very simple by using the Bresenham algorithm to extract pixels along the corresponding epipolar line. The rectified images are guaranteed to be effective for all possible camera motions, large or small. The loss of pixel information along the epipolar lines is minimized, and the size of rectified image is much smaller. Furthermore, it never splits the image and the connected regions will stay connected, even if the epipole locates inside an image. The effectiveness of the method is verified by an extensive set of real experiments. It shows that much more accurate matches of feature points can be obtained for a pair of images after the proposed rectification.6. A novel metrology without using any contacted tool is pr更多还原