【作者】 赵连军；

【导师】 刘恩海； 张文明；

【作者基本信息】 中国科学院研究生院（光电技术研究所） ， 信号与信息处理， 2014， 博士

【摘要】 单目视觉位置姿态测量系统是视觉测量系统的重要组成部分。与双目视觉相比，它具有结构简单、测量精度高等优点，被交会对接、机器人、无人机、自动检测设备等广泛采用。为解决视觉位姿测量在空间交会对接工程应用中存在的问题，本课题开展了基于目标特征的单目视觉姿态测量系统的研究，包括基于合作目标的位姿测量和基于非合作目标的位姿测量两种。国内外交会对接中通常采用基于合作目标的位姿测量系统，测量系统主要由相机、合作目标、特征提取和位姿解算算法、激光照明和控制系统等部分组成，其中特征提取和位姿解算算法两部分是系统的核心。分析了基于合作目标位姿测量中特征提取的原理，试验验证了特征提取原理的可行性；建立了合作目标位姿测量精度模型，分析了三点位姿测量中影响测量精度的6个主要因素，包括测量距离、特征点之间的位置关系、图像上特征点的提取精度、相机的量化误差、相机内参数标定误差、位姿解算算法，分析了各因素对总体精度的影响方式和影响程度。利用重投影的方法验证了测量系统的精度，在2.4米处测量系统重投影的像素误差为（0.05，0.05），对应的位置误差约为0.12mm。合作目标提高了测量系统的稳定性和精度，但是限制了测量系统的应用范围，例如在对没有预先安装合作目标的飞行器进行在轨维护和升级等。为了将单目视觉位姿测量系统应用到上述情况中，本课题开展了基于非合作目标的位姿测量系统。测量系统的开发难度主要是特征提取和位姿解算算法两个方面。非合作目标位姿测量中没有激光照明系统配合提取特征点，因此特征提取难度加大、提取精度降低。本文根据使用目标的特性提出了利用全局信息提取目标特征的方法，利用特征拟合等方法提取目标的隐含信息，实现了图像中特征和目标上特征的匹配。非合作目标上的特征点没有按照测量需求排列，而且能够提取的特征类型不固定，所以位姿解算方法不固定，需要根据不同的情况设计相应的位姿解算算法。根据课题中使用的目标上能够提取的特征，提出了利用目标的圆心和半径等信息计算相对位姿的算法，然后分析了算法的误差来源，提出了误差修正方法。在理论分析的基础上实验验证了算法的正确性和稳定性，利用重投影的方法检验了算法的精度，在2.8米处测量系统重投影的像素误差为（0.4795，0.5606），对应的位置误差为1.48mm。精度分析和误差分配是测量系统开发的重要组成部分，精度分析能够指导系统设计、误差分析、误差分配等环节。利用更高精度的测量仪器分析视觉测量系统的误差，试验中使用全站仪标定相机参数和分析测量系统精度。通过试验分析得到基于合作目标的位姿测量在测量距离为2.4米左右的位置测量精度为2.6‰，三个坐标轴上的姿态角测量精度优于0.2°；基于非合作目标的位姿测量在相对距离为2.8米左右的位置测量精度为6.0‰,姿态角测量精度优于0.3°。本课题完成了基于单目视觉测量的框架性理论和实验验证，利用全站仪进行了相机标定和测量精度分析工作：建立了合作目标位置测量的误差模型，能够指导系统设计和误差分配；提出了提取图像中非合作目标特征的方法和利用非合作目标特征解算位姿的方法，分析了方法的误差、完善了方案，试验验证了方法的正确性和稳定性；提出了测量系统精度分析方案，标定了相机的内外参数，分析了合作目标和非合作目标测量系统的精度。更多还原

【Abstract】 Mono-vision measuring system is an essential consisting part of visionmeasuring region. Mono-vision measuring system is widely adopted by RVD(Rendezvous and Docking), auto detection machine for its simplicity, highprecision. In order to extend the application of mono-vision system to robotic arm,robotic hand, RVD, the research is carried out based on both cooperative target andnon-cooperative target.Measuring system based on cooperative target is consisted by camera,cooperative target, feature extraction algorithm, position and orientation resolvingalgorithm, laser system and control system, among all of which, feature extractionand pose resolving algorithms are the most important parts. Image contains target issubtracted by background image without the target when feature points areextracted during RVD is in operation. Pose resolving algorithm by three featurepoints is regularly adopted by cooperative target measuring system. Precision ofthree points algorithm is influenced mainly by6aspects, such as measurementdistant, feature points character, and precision of feature point extraction.There is no cooperative target in many working conditions, for example, Hubblespace telescope maintenance on orbit, size measurement of workpiece, and licenseplate number detection. In order to extend pose measuring system to thoseconditions that pose measurement system based on non-cooperative target isdeveloped. The measuring system is consisted by camera, non-cooperative target,feature extraction algorithm, and pose resolving algorithm. Dimensionalinformation of target could not be reconstructed by mono-vision system, so thatdimensions of target should be known information. While precision of featureextraction is the most influencing factor, extraction of feature points onnon-cooperative is considerably tougher and less precise than cooperative target.Pose resolving algorithm is another intractable problem for that feature type is notconsistent. Feature points, feature lines, curves, and regions are used for posecalculation, so that the pose resolving algorithms should be capable to get poseinformation from those features. Usually, the pose resolving algorithm is varying indifferent conditions.Precision analysis and accuracy distribution are two important parts after measuring system is developed. Camera calibration, which is a preliminaryprocedure for precision analysis and accuracy distribution, is consisted by intrinsicand extrinsic parameters calibration. Intrinsic parameter is composed by focallength, center position of image, size of pixel on CCD or CMOS, and extrinsicparameter is composed by position and orientation relationship between cameracoordinate and reference coordinate. Total station is adopted for the cameracalibration and precision analysis, three corner cubes are imposed on both the targetand camera frame as the target of total station. Position accuracy of cooperativetarget is2.6‰when the measurement distance is about2.4meters, and the angleaccuracy is better than0.2°. Position accuracy of non-cooperative targetmeasurement is6.0‰when the measurement distance is about2.8meters, andangle accuracy is better than0.3°.Camera model, camera calibration model are included as preliminaryprocedures for pose measurement. Pose calculation algorithms by three featurepoints, four points, and N points are deduced, and feature points extractionmechanism for cooperative target is presented, and image processing procedure arealso analyzed. At the same time, feature extraction methods for points, lines, andcurves, and methods for finding correspondence between features extracted fromimages and on the model are presented for non-cooperative target. Methods to poseresolving for non-cooperative target by varying feature types are developed. Totalstation is adopted for camera calibration and accuracy analysis. The frame theoriesof mono-vision measurement are deduced, experiments are carried out, andprecision analysis is presented.更多还原