【作者】 苑云；

【导师】 于起峰； 张小虎；

【作者基本信息】 国防科学技术大学 ， 航空宇航科学与技术， 2013， 博士

【摘要】 随着科技的发展，越来越多的大型结构出现在人们的生产生活中，诸如桥梁、隧道、飞机、飞艇、风电叶片、天线等。全场、高精度、动态测量大型结构的形貌、变形，即形变参数，是其动态性能实验和质量监测的基本要求。摄像测量具有非接触、高精度、可测点多、实时动态测量等特点，在大型结构的形变测量方面有独特的优势。但是，由于大型结构的形变测量同时要求大尺度和高精度，大尺度要求像机具有大视场，高精度需要有高空间分辨率，双像机摄像测量方法在这两者之间存在矛盾。为此，本文采用组网摄像测量方法，研究了该方法在大型结构形变测量应用中亟需解决的四个关键问题：像机组网优化、高精度像机标定、实时三维交会和快速平差优化。本文取得的主要成果如下：1.提出了基于改进遗传算法的摄像测量网络设计方法。待测大型结构复杂，组网摄像测量中像机数目和类型多，环境约束多，仅凭经验布置像机，不能保证像机数量已确定情形下的测量精度最优。因此，组网摄像测量首要解决的问题是在给定待测模型、像机类型和数量、测量环境约束等条件下，优化设计满足精度要求的布置形式，即摄像测量网络设计。由于摄像测量网络设计具有多参数、多约束、运算量大等特点，本文使用改进的遗传算法随机寻找全局最优解。本文算法中考虑了可见性和可测性判断、分块的反向误差精度估算等因素，能够得到满足约束条件和测量精度的组网配置结果，为实施现场组网测量提供依据。2.提出了组网摄像测量中的一系列像机标定方法。由于多像机组网提供了足够的多视图几何约束，组网摄像测量系统可用两视图自标定结果作为初值，统一平差优化所有像机参数，实现系统的高精度标定。此外，为解决像机标定中各参数耦合的问题，研究了逐级独立的像机标定；为解决待测区域难以布置大量控制点的问题，研究了较为相应的大视场像机标定；作为固定像机的有益补充，经纬像机能够全场高精度测量，本文研究了中远场测量中经纬像机的标定。2.1提出了从焦距简便解析解自标定变焦两视图的方法。该方法仅需要已知像机的长宽比和零扭曲这两种内在约束，即可由同名像点得到等效焦距的更简单的解析表达式，进而标定像机，并对目标进行欧氏重构。2.2提出了基于位姿估计变化量逐级标定像机的方法。该方法首先基于位姿估计变化量标定等效焦距，然后逐级标定其他参数，解决了像机内、外参数之间耦合的问题，各参数值与其物理意义相符，可独立使用。2.3提出了基于无穷单应的大视场像机标定方法。该方法最少只需要四个非共线控制点和像机粗略的位置即可求解无穷单应，并且提出了一种坐标变换方法以保证线性求解和优化无穷单应时的稳定性。2.4提出了准同心广义经纬像机的成像模型及高精度标定方法。对于广义经纬像机的标定，在近场及单幅图像内控制点数目比较多时，可使用手眼标定等方法，但在中远场标定与测量时这些方法不适用。为方便中远场测量的经纬像机的标定，本文假设像机与旋转平台准同心，提出了准同心广义经纬像机的成像模型及高精度标定方法。3.提出了最小化空间距离的线性交会方法。交会测量时平差物点的迭代方法虽然能够保证精度，但是运算量大，不能保证实时。本文所提方法为线性方法，交会方程的物理意义为最小化物点到各像点反投影射线的距离平方和，只需一步解算得到物点坐标，精度与平差物点的迭代方法相当。4.提出了两种平差方法：共内参多视图物点形式的稀疏LM（Levenberg-Marquardt）光束法间接平差；附有限制条件的稀疏LM光束法间接平差。光束法平差是提高测量精度的必要步骤，稀疏分块平差能显著提高平差速度，且考虑既有约束的平差模型才更加符合真实物理意义。本文提出的第一种平差方法利用了多视图共内参的约束，能够同时优化物点，适用于在控制点精度不高情形下的像机标定。第二种方法能够灵活设定参数是否需要优化，能够添加各种内在约束、场景约束和运动约束，但仍然保留了稀疏分块结构，适用于存在各种约束的大型结构形变测量。利用上述研究成果，设计和研制了大型结构形变组网摄像测量系统，成功应用于风电叶片、机翼和飞艇等的形变测量。该系统能够以较优的摄像测量组网配置，测量大型结构的整体变形和各个局部区域变形，适用于场景比较单一的区域中大视场像机的标定和测量，也适用于待测区域内无任何控制信息的测量。组网摄像测量方法为大型结构形变参数的实时/准实时、动态、高精度测量提供了有力的理论和技术支持，有效拓展了传统摄像测量的应用范围，增强了摄像测量在宏观测量方面的工程实用性，有广泛应用前景。更多还原

【Abstract】 With the development of science and technology, more and more large-scalestructures come into our lives, such as bridge, tunnel, airplane, airship, wind turbineblade and antenna, etc. It’s a basic requirement of corresponding dynamic performanceexperiment and quality monitoring to measure the shape and deformation of theselarge-scale structures universally and dynamically with high precision.Compared with other measuring means, videometrics has a lot of advantages in thedeformation measurement of large-scale structure, it can conduct non-contact dynamicmeasurement with high precision, and can handle large set of points simultaneously. Butdeformation measurement of large-scale structure has requirement on measuring scaleand precision at the same time, that is the large field of view and high spatial resolutionrespectively, which brings contradiction for dual-camera videometrics. To cope with thisproblem, networked videometrics is proposed in this dissertation, and fourcorresponding critical problems are studied: camera network optimization, cameracalibration with high precision, real-time3D intersection and rapid bundle adjustment.Main contributions of this dissertation are as follows:1. Videometric network design based on improved genetic algorithm is proposed.The structure to be measured is quite complex, besides, cameras to be used innetworked videometrics are many and versatile, and there are a lot of environmentalrestrictions, so the optimal measuring accuracy given the amount of camera cannot beguaranteed only by experience. Thus, the first problem that has to be solved is todetermine the layout of cameras, given the model to be measured, along with the typeand the amount of camera and environmental restrictions, in order to achieve therequired measuring accuracy, and this procedure is so called videometric networkdesign.For videometric network design is a computationally exhausting nonlinearprocedure with multiple parameters and constraints, improved genetic algorithm isadopted in this dissertation to search for global optimum randomly. Kinds of factors,such as visibility, measurability and block reverse error precision estimation, etc. areconsidered in the algorithm, design results satisfying the constraining conditions andmeasuring accuracy are acquired, which can provide reliable reference for conductingfield networked measurement.2. A series of camera calibration methods is proposed for networked videometrics.Since multiple cameras are used in networked videometrics, there are adequatemultiple view geometric constraints. Self calibration result based on two views can beadopted as initial values for networked videometrics system, followed by bundleadjustment to obtain calibration result with high precision. Besides, to cope withcoupling among each parameter in camera calibration, stepwise camera calibrationmethod is studied. Sometimes, it is difficult to locate large set of control points in the area to be measured, to cope with this, flexible calibration method for large field ofview camera is studied. As a beneficial supplement to fixed camera, Theodolite-cameracan conduct full field of view measurement with high precision, correspondingcalibration method when using theodolite-camera for middle-field and far-fieldmeasurement is studied.2.1A camera self calibration method for two-view with varying focal length, whichis originated from simplified analytical solution of equivalent focal length, is proposed.Given only aspect ratio and zero-skew, this method can get the simplifiedanalytical solution of equivalent focal length from corresponding points of two images,and then cameras are calibrated and Euclidean structure is reconstructed further.2.2A stepwise camera calibration method based on pose estimation variation isproposed.In this method, first equivalent focal length is determined based on pose estimationvariation, and then other parameters are determined step by step. This method caneffectively avoid coupling among each camera parameter, in which way the value ofeach parameter can match its own physical meaning pretty well, and can be usedindependently.2.3A calibration method for large field of view camera based on infinitehomography is proposed.This method only needs approximate position of the camera and at least fournon-collinear control points to compute the infinite homography. A coordinatetransformation method is also proposed to ensure stability of the linear solution andoptimization process of the infinite homography.2.4Imaging model and high precision calibration method for quasi-concentricgeneral theodolite-camera is proposed.As for general theodolite-camera, it can be calibrated by hand-eye calibrationmethod when used for near-field measurement, if there are adequate control points in asingle image, which is not suitable for middle-field and far-field use. In order toexpediently calibrate the theodolite-camera used for middle-field and far-fieldmeasurement, we suppose the optical center of camera is quasi-concentric with therotation center of rotation frame, and its corresponding imaging model and highprecision calibration method is proposed.3. A linear intersection method, which minimizes the spatial distance, is proposed.As for the intersection, although the iterative method can guarantee accuracy, it iscomputationally expensive and cannot guarantee efficiency. Proposed method is linear,in which the geometric meanings of intersection equation is to minimize the sum ofsquares of the spatial distances from the object point to the back-projected radial of itseach corresponding image points. Only one-step resolvent is demanded, and themeasuring accuracy is very approximate to the ones of iterative method.4. Two bundle adjustment methods are proposed. One is the sparse LM indirect bundle adjustment with the form as mutual internal parameter, multiple view and objectpoints. Another is the sparse LM indirect bundle adjustment with restricted condition.Bundle adjustment is an imperative procedure to refine measuring results, in whichthe sparse block mechanism can speed the adjustment procedure up significantly, andconsidering existing restricted condition should be in accordance with the true physicalmeanings. The first method is applicable for the calibration of cameras with the samefixed internal parameters and can optimize the positions of the reference pointssimultaneously, it is suitable for camera calibration under situations in which theaccuracy of reference points is not good enough. The second method can easily set theparameters whether be adjusted or not, and can add various constraints such as intrinsicconstraint, scene constraint and motion constraint. Besides, it still preserves sparseblock structure, so it is applicable for the deformation measurement of large-scalestructure which has various constraints.Utilizing above research achievements, we have designed and developed apractical networked videometrics system for the deformation measurement oflarge-scale structure, and applied it to practical deformation measurement of windturbine blade, airplane wing and airship, etc. This system can measure the integraldeformation and all of the local deformation. The system is applicable for thecalibration and measurement of large field of view camera in simplex scene, and for themeasurement of regions without any control information for calibration.Networked videometrics method provides a powerful theoretical and technicalsupport for real-time/quasi real-time, dynamic and high precision measurement oflarge-scale structure deformation. It effectively expands the application range oftraditional videometrics, and enhances the engineering practicality in the aspect ofmacroscopic measurement of videometrics, and has a wide and important applicationalprospect.更多还原