【作者】 周拥军；

【导师】 寇新建；

【作者基本信息】 上海交通大学 ， 固体力学， 2007， 博士

【摘要】 利用视觉测量或数字近景摄影测量技术进行高精度三维测量与重建是当前计算机视觉和数字摄影测量领域的重要研究方向,在大型土木工程结构的变形测量中,采用传统的千分表或其它接触式传感器的方法往往不能满足要求,而随着高分辨率CCD数码相机的广泛应用,运用普通数码相机实现非接触的光学三维测量日益广泛。传统的数字近景摄影测量采用专业量测相机或已检校的普通数码相机,而在某些场合难以实现,论文研究了无足够物方控制条件下运用未检校CCD相机实现三维测量的基本理论和方法,并将其应用于结构试验变形监测中。1.论文比较了摄影测量领域的共线方程和计算机视觉领域的线性或非线性针孔模型的联系与区别。分析和比较了经典摄影测量中的空间交会方法、直接线性变换方法、相对定向-绝对定向、光束法平差等解析处理方法的理论模型、解算方法以及应用场合;总结归纳了外极线约束、基本矩阵、本质矩阵、单应矩阵的估计方法及其与相机内、外方位元素的关系。对于无物方控制点的未检校影像,基本矩阵是唯一能从立体像对中恢复的元素,本文提出了附约束条件最小二乘稳健估计法;2.提出了适用于数字近景摄影测量的点特征的提取和匹配策略并分析了所能达到的精度,即先对可编码人工标志进行自动检测并用基于矩和曲率保持方法提取子像素边界后,用最小二乘拟合方法对标志中心精确定位,然后恢复外极线,采用基于外极线约束的最小二乘匹配,并辅之以必要的人工量测;3.针对经典摄影测量采用欧拉角表示旋转矩阵的不足,本文研究了采用单位四元数表示旋转矩阵的方法,推导出了单位四元数表示方法与欧拉角表示方法、罗德里格矩阵表示方法之间的转换关系。并将单位四元数表示的旋转矩阵应用于相对定向、绝对定向、光束法平差中。改进了相对定向的PH算法,使得使用该方法进行相对定向不需要初值时也能实现全局收敛,采用基于单位四元数的绝对定向的闭合解,不需要初值也能得到定向参数的最小二乘解。4.研究了无任何物方信息情况下双像、多像三维度量重建的方法,提出基于单位四元数的附约束条件一维搜索迭代法进行广义相对定向的新方法,和现有的直接解法相比,该方法极大地提高了广义相对定向的精度和稳健性,使得利用两张未检校影像实现度量重建在绝大多数条件下可行;改进了计算机视觉领域的多像层次重建方法并应用于立体摄影测量内、外方位元素的初值估计之中,有效地解决了摄影测量中无物方控制信息的相机方位元素的初值估计问题;5.研究了相机单站纯旋转的理论模型,提出了采用基于显参数的最小二乘估计方法代替分步估计方法,该方法极大地提高了自检校的精度和稳定性。建立了考虑旋转偏心条件下,相机做Pan-Tilt-Zoom运动的投影模型以及三维测量方法。该方法包括同站相对定向、基片扩充、异站定向、绝对定向四个步骤。其中同站相对定向理论是对经典摄影测量理论的创新,同站定向在理论上减少了定向参数,只需要少量的同名点即可完成,有利于提高定向精度。可通过架设于普通脚架上的数码相机的纯旋转实现摄影测量,有助于改善大型构件摄影测量的构像条件;6.自检校光束法平差是理论最严密、解算精度最高的摄影测量解析处理方法,在获取内、外方位元素的初值后,高精度的测量成果需要用光束法平差求精。论文研究了附基准条件自检校光束法平差的统一模型,这一模型既适合于经典摄影测量控制网,也适用于摄影测量自由网,对于基于摄影测量的变形监测具有重要意义。对于三维变形监测数据,变形分析和预测是一项主要内容,论文给出了基于误差椭球的假设检验方法以及基于小波变换的去噪方法;特别是后者,对于分析观测周期长、位移量小的观测数据序列具有比常规处理方法更好的效果;7.基于数字摄影测量的位移测量方法具有非接触、可测量的点多、速度快的特点,具有常规的位移测量方法不可替代的作用,以结构模型试验变形监测为应用背景,基于Matlab软件开发了一套实用的计算程序,并以高速公路路堑边坡的模型试验和玻璃钢夹砂顶管的力学性能试验为例介绍了摄影测量法在位移测量中的应用。更多还原

【Abstract】 High precision three-dimensional(3D) reconstruction and vision metrology are one of the most active research areas in Computer Vision and Digital Close-range Photogrammetry society. In the large-size model experiment of civil engineering and mechanics,displacement or deformation measurement is indispensable, dial indicators and other displacement sensors are the traditional tools used for displacement measurement, but it is a contacted approaches and only with the capacity of one-dimensional measurement, so it can not work well in most cases. Photogrammetry is a non-contacted, high accuracy and automation techniques for 3D measurement which could be utilized as an effective tools in these fields, traditional photogrammetry relied on professional metric camera and other expensive facilities, But nowadays, with the development of high resolution Charged Coupled Device(CCD) solid state camera, it is widely used in close-range photogrammetry society instead of professional cameras, so the new theories and approaches are needed. Under this background, this paper studies on the uncalibrated CCD camera based digital close-range photogrammetry for 3D structure deformation measurement without enough information about object space. The contents of this paper are given as follows:1. Camera models used in computer vision and photogrammetry society are compared in thesis, The linear and nonlinear pinhole camera in computer are the same as the collinear equation with or without lens distortion. The mathematical model, resolution algorithms, applicable situations of the analytical approaches on photogrammetry such as direct linear transform, spatial resection and intersection, relation orientation and orientation and bundle adjustment are introduced, and the epipolar constraints, fundamental matrix, essential matrix and homography matrix on computer vision are introduced as well. As the fundamental matrix is the unique information which can be recovered from the uncalibrated image sequences without any prior information of object space, a new estimation algorithm named constrained robust least square method for fundamental matrix is given.2. A practical approach for point feature extraction, precise location and matching is proposed for close-range image sequences. With this approaches, the artifical coded targets are automatic detected, and the sub-pixel edges are accurately extracted with moment and curve preserving method and the accurate center are obtained with least square fitting, then the epipolar line of any two overlapped images are recovered, the LSM matching with epipolar constraints is applied for accurate feature location, and manual measurement on screen are designed if necessary. The error sources and the theoretic precisions of each measurement approach are analyzed;3. Unit quaternion are proposed for representing rotation matrix instead of traditional Euler angles, and the transformations of each representing approaches are also given. The full estimation approaches on relative orientation, absolute orientation and bundle adjustment are proposed. An improvement on PH algorithm in relative orientation are made, with this algorithm, the iterative procedures could be global convergent without any initial values. The unit-quaternion based closed-form solution for absolute orientation is a least square solution without initial values.4. 3D metric reconstruction methods with two or multiple uncalibrated views without any scene knowledge are studied, a practical quaternion based iterative algorithm for general relative orientation are proposed for two-view metric reconstruction, compared with the existed direct approaches, robustness and accuracy are improved. And a stratification 3D reconstruction method in computer are introduced to photogrammtry for interior and exterior parameters estimation, which is very helpful to initial orientation parameters estimation;5. The single station pure rotation based self-calibration approaches are studied, compared with the existed two- step estimation method, a LSM method based on explicit parameters model are proposed for self-calibration, which could improve the stability and accuracy greatly. The projection model under the assumption of camera mounted on a tripod undergoes pan-tilt-zoom(PTZ) motion with rotation eccentricity are given, and the 3D measurement of PTZ camera includes single-station relative orientation, base image extending, two-station relative orientation and absolute orientation. Where single-station relative orientation is an innovation for traditional photogrammetry, it can achieve more accuracy and need less point correspondences, and improve the geometry configuration for image sequences, which is very useful for large structures 3D measurement;6. Self-calibration bundle adjustment method is the most precise analytical method for photogrammetry and high precision 3D measurement usually uses this method for refinement when initial values are known. A unified self-calibration bundle adjustment model with datum constraints is given, which is not only appropriate for traditional photogrammetry network adjustment but for free-network. Deformation data analysis is an important task for deformation monitoring, an error ellipsoid based hypothesis test method and a wavelet based denoising method are proposed for deformation data analysis;7. Photogrammetry based 3D measurement is a non-contacted, high-efficient technique for displacement measurement, which is indispensable for some special situations.Under the background of structure displacement measurement, a Matlab program is developed. And two experiments are given for verifying the results of displacement measurements.更多还原