【作者】 李莉；

【导师】 曾庆双；

【作者基本信息】 哈尔滨工业大学 ， 控制科学与工程， 2009， 博士

【摘要】 计算机视觉技术在工业、医疗、航空航天等领域具有广泛的应用前景,论文在总结国内外计算机视觉领域的研究现状和发展趋势的基础上,系统、深入地研究了摄像机标定、图像配准、视觉跟踪和视觉测量等计算机视觉系统中的几个关键问题,并将理论研究成果应用于相应的计算机视觉任务中,验证了本文所提出的方法的可行性和有效性。首先,研究了标定帧清晰度不理想情况下的摄像机标定问题。首先分析了基于平面标靶的两步标定法的基本原理,给出了衡量标定结果误差的图像畸变测度计算方法,将本文中不理想条件下的标定问题转化为主控制点的优化问题,采用粒子群优化法实现优化过程。仿真结果验证了该算法的有效性,该标定算法在论文第四章的视觉跟踪任务中得到应用,验证了该算法的工程实践价值。其次,研究了一类幅值非恒定的图像变换模型的图像配准问题。通过引入幅值变化系数来表征图像灰度差异,建立了幅值非恒定的图像变换模型。针对幅值非恒定的平移变换模型,证明了扩展差值函数有零点与平移量之间存在线性关系,提出了基于扩展差值函数的配准算法,仿真结果表明该算法具有较高的抗噪性和较低的计算复杂度。针对幅值非恒定的旋转变换模型,证明了待配准图像与基准图像在一定角度上的Radon变换矢量线性相关,提出了基于Radon变换矢量匹配的配准算法。针对幅值非恒定的刚体变换模型,提出了基于角度差值函数和径向差值函数的配准算法,该算法消除了冗余系数的负面影响,仿真结果表明该算法在保证计算精度的同时降低了计算的复杂度。该算法应用于论文第四章和第五章的视觉任务,验证了其在工程实践中的可行性和有效性。然后,研究了单目和双目视觉跟踪问题。针对单目视觉跟踪问题,提出了改进的均值漂移跟踪算法,建立了均值漂移跟踪的模型更新机制,解决了核函数窗宽的自动选取问题,降低了均值漂移跟踪算法对背景颜色的要求,并提高了对变外观目标的跟踪能力。针对双目视觉跟踪问题,采用图像拼接技术实现了对多视频源图像的信息整合,将多台摄像机拍摄的视场图像拼接成全局视场图像,扩大了系统监控范围,实验结果验证了该算法的可行性。最后,研究了非合作目标的视觉位姿测量问题。设计了两段式非合作目标位姿测量方法,采用运动式立体视觉测量技术将非合作目标转化为合作目标,在此基础上使用P3P位姿测量法计算非合作目标的位姿运动参数,为非合作目标视觉测量提供了崭新的解决思路。该方法在空间交会对接地面仿真系统的视觉测量子系统中进行了试验,现场实验结果验证了本文视觉位姿测量方法的可行性和有效性。更多还原

【Abstract】 Computer vision technologies have been widely used in industries, medical, aerospace and other fields, based on concluding the state of the art and future trends in the field of computer vision system, this dissertation systematically studied issues on camera calibration, image registration, visual tracking and visual measurement in-depth, and applied the theoretical research results to the corresponding computer vision tasks, which verify the feasibilities of the proposed algorithms.(1) The camera calibration issue is studied for non-ideal circumstances when the calibration object image clarity is not sufficient. Firstly the basic principles of planar-based two-step calibration algorithm are analyzed, then the image distortion measurement calculation method is given for measuring error of the proposed calibration algorithm, The calibration problem under non-ideal conditions is transformed into the optimization problem of the main control points, ultimately the particle swarm search strategy is utilized to accomplish the optimization process. Experimental results validate the effectiveness of the proposed algorithm, while the method is used in chapter 4 for image distortion correction part in the visual tracking task, demonstrates its practical value of the algorithm.(2)Image registration problem is studied for a class of non-constant amplitude image transformation model. Through the introduction of the amplitude difference coefficient describing the image gray level variation, the non-constant amplitude image transformation model is established. For non-constant amplitude translation model, the linear relationship between one zeros of the extended difference function and 1D translation is proved, and the extended difference function based image registration method is proposed. Experimental results show that the method is applicable to variant gray level images with good noise robustness and low complexity. For the non-constant amplitude rotation model, it is proved that Radon transform vectors can be taken as image feature for matching, and then a registration method is proposed based on Radon transform vectors matching, which utilizing multi scale searching strategy to search for the rotation angle. For non-constant amplitude rigid body transformation model, it is proved that the angular difference function and radial difference function can be used to solve the image rotation angle and scaling factor, and then a registration method is proposed based on the angular difference function and radial difference function. Simulation results show that the algorithm can effectively reduce the computing complexity, while achieving same accuracy with classic phase-correlation based method. The registration methods are verified effectively for vision tasks in Chapter 4 and Chapter 5.(3)Visual tracking issue is studied. For monocular visual tracking issue, an improved mean-shift tracking algorithm is proposed, using rigid-body model based image registration method to align images in tracking window, in this way establishes an auto-updating mechanism for tracking model. The method resolves the target missing problem resulting from similar color between foreground and background, target scaling changes, and target appears changes. For binocular tracking issue, a global view tracking algorithm is proposed based on image mosaicing technology. Experimental results show that the algorithm is effective.(4)The non-cooperative target vision pose measurement issue is studied, a two-stage non-cooperative target position and attitude measurement method is proposed, using motion stereo vision measurement technology to transform non-cooperative target into cooperative target, after that, P3P pose measurement method is used to estimate non-cooperative target position and attitude parameters. Motion stereo vision measurement technology utilizes monocular camera to simulate binocular stereo visual system to measure the target’s features geometry size, and then to determine geometric structure of features, as the cooperation of non-cooperative targets. This method has been successfully applied to the space rendezvous and docking simulation system as visual measurement subsystem, the experiment results validate the feasibility of the proposed method.更多还原