【作者】 臧林然；

【导师】 王晓明；

【作者基本信息】 大连理工大学 ， 机械电子工程， 2003， 硕士

【摘要】 计算机视觉是实现模式的自动识别即以计算机完成对视觉信息处理的科学，是基础研究和应用研究中重大的挑战之一。而图像匹配由于其涉及的问题众多，是计算机视觉中最困难、最关键的一步。为了认识影响3D重建精度的来源和规律，在计算机视觉系统中引入三维重建的不确定性描述是必要的。 本文探索性地提出了一种基于梯度场相似性和邻域膨胀的快速图像匹配算法。用SUSAN法检测到我们感兴趣的特征一角点，改进的Scott和Longuet-Higgins算法完成角点的匹配，进而进行基础矩阵估计。从一点的某邻域内梯度场相似性出发，利用极线约束完成全象素匹配。提出的邻域膨胀方法能有效处理多候选点的问题，与传统的利用连续性约束、匹配强度算法等相比能明显加快匹配速度，自适应邻域的匹配策略能够保证较高的精度和可靠性，算法受光照条件的影响较小且适合不同尺度图像。 方法的多样性及采用的分析工具的不同，建立统一的误差分析模型并不现实。由于得到基于误差传播理论的重建不确定性显式公式并没有直观性，本文在推导出三维重建扰动分析模型的基础上，应用多元分析的统计方法研究了图像量化误差、匹配误差、标定误差等对重建精度的影响。采用计算机仿真图像进行实验，向重建模型中输入高斯噪声进行扰动分析，这样有利于对不确定性的评定。对三维重建点伸展不确定性的可视化也进行了探讨。本文给出的扰动模型和多元分析的方法具有更大的通用性。 对提出的匹配算法及重建不确定性研究的理论与方法，本文给出了其软件实现及实例分析结果。介绍了模块的整体设计并给出几个关键问题的程序分析。实验验证了匹配算法的有效性并得出了重建不确定性研究的若干重要结论。更多还原

【Abstract】 Computer vision is the science of implementation of pattern recognition automatically in which computer finishes information processing and which is really a tremendous challenge during the process the fundamental and practical research of mankind.Image matching is most difficult and crucial in computer vision for which involves numerous problems. To understand factors and rules that influence the precision of reconstruction, to introduce 3D reconstruction uncertainty analysis is necessary. This paper presents the research of problems mentioned above.A fast images matching algorithm is proposed in this paper, which is based on gradient similarity and neighbor expansion. Interesting features (corners) are detected using SUSAN method and they are matched with improved Scott and Longuet-Higgins algorithm, the fundamental matrix is estimated after that. From the gradient field similarity of the two points, all pixels can be matched under the constraint of epipolar geometry. The neighbor expansion method and self-adapted neighbor selection strategy can handle the multi-candidate problem effectively and robust, rapid enough to be implemented unlike methods with continuity constraints and matching intensity algorithm. The algorithm is slightly effected by lighting environment and suitable for multi-scale images.Uniform precision analysis model is hard to achieve for large numbers of methods and tools used. Since the explicit uncertainty equation of 3D reconstruction achieved by error propagation theory is not intuitionistic, this paper presents a perturbation analysis model of 3D reconstruction and studies reconstruction precision affected by image digitalization error, matching error and calibration error using multidimensional analysis of statistic method. Gaussian noise is added to the reconstruction model for perturbation analysis using synthetic images, thus is helpful for uncertainty evaluation. Finally, expanded uncertainty of the reconstructed points is visualized. Perturbation analysis model and multidimensional analysis are more universal.Software of the matching algorithm and reconstruction uncertainty method is implemented and experiment results are given. Gross module design is introduced and details are presented for key problems. Experiments tested the effectiveness of the matching algorithm and important conclusions of reconstruction uncertainty problem are’ brought.更多还原