【作者】 郁钱；

【导师】 须文波；

【作者基本信息】 江南大学 ， 计算机应用技术， 2011， 硕士

【摘要】 计算机视觉的目标是通过感知的图像理解世界中的各种物体。需要理解的物体的属性信息众多,其中最重要的信息是物体的整体三维结构,所以基于视觉图像的三维重构自从计算机视觉的诞生以来就成为计算机视觉研究的热点和重点。经过将近三十年的研究,此问题已经得到部分解决,并促进了计算机视觉一门分支—计算机多视图几何的诞生。基于视觉图像的三维重构涉及到四种关键技术,包括摄像机标定、特征提取和匹配、运动估计和结构计算。论文从已有的研究成果的基础上,做了以下五个方面的工作:(1)为了克服传统优化算法的缺点,提高摄像机标定的精度,第3章首次将量子粒子群优化算法应用到摄像机标定中。首先,该方法运用传统的线性方法获得初始值,然后使用QPSO优化初始值,得到一个比较优秀的值。实验数据表明,基于QPSO的摄像机标定的平均反投影误差小于一个像素,是一种可行的方法,且与智能优化算法PSO相比,基于QPSO的摄像机标定具有更小的误差。(2)论文指出如果两个平面间存在四组点对应,则这两个平面的笛卡尔坐标系之间存在一个非奇异线性变换,即平面单应。这就是射影几何中著名的莫比乌斯定理。2.5节从射影几何的基本定义出发,严格证明了莫比乌斯定理,并给出了计算平面单应的方法。(3)运动估计是基于视觉图像的三维重构的核心问题,所谓的运动估计就是从拍摄得到的多幅图像中计算出摄像机之间相对位置的过程。Hartley首次给出了从本质矩阵恢复运动的一种估计方法。第5章对该方法进行了深入的研究,首次对Hartley方法提出了一种简单易理解的证明过程。(4)一般的三维重建系统利用角点等作为特征点进行匹配,这些匹配容易出现较高的错误匹配率。为提高匹配的准确度,第7章构建的实验系统用尺度空间中不变点作为特征点,使用SIFT算法提取和匹配这些点。实验效果显示使用该方法是有效的,具有一定的应用价值。(5)第4章提出了一种直观且简单的不变特征点提取的统一思想框架,统一了SIFT、SURF以及HARRI等特征点提取算法,指明了不变特征点提取的研究方向。更多还原

【Abstract】 The goal of Computer Vision is to make decisions about all kinds of objects in the world from the sensed images. These images fulfill numerous information about their properties in which the three dimentional structure is the most noteworthy, so Three-Dimensional Reconstruction from Perspective Views is an important and hot point since the birth of computer vision. The research topic has been partly resolved well and brought a new decipline - multiply views geometry - of Computer Vision. There are just four key technologies in respect to this research topic, including camera calibration, feature extraction and matching, motion estimation, and structure computation. Based on the previous research, this thesis has done some work followed from five aspects:(1)The Quantum-Behaved Particle Swarm Optimization has been firstly applied to camera calibration, in chapter 3, in order to improve the accuracy and overcome the drawbacks of traditional optimization algorithm. Firstly, this method uses the traditional linear method to achieve the initial value, and then optimizes the initial value with QPSO. Experimental data shows that camera calibration based on QPSO has less average back-projection error than a pixel and is an effective and reliable method. Experiment also shows that this approach has lower error than the one based on PSO.(2)The thesis points out that if there are four corresponding pairs of points between two planes, there will be existing one non-singular linear transformation, ie homography. This is the famous Mobius theorem in projective geometry. Section 2.5 proves the Mobius theorem from the basic definition of projective geometry and gives the homography calculation method.(3)Motion Estimation is the core issue in Three-Dimensional Reconstruction from Perspective Views, which is to calculate the relative positions among cameras from multiple images taken of one object from different viewpoints. One outstanging method has firstly appeared in Hartley[27][28] for Estimating Motion from Essential Matrix. Chapter 5 does some deep work on this method and brings out a new and easy proof to it.(4)The general systems of Three-Dimensional Reconstruction often use corner as feature points for matching, but the rate of matching is prone to be higher. To improve the accuracy of matching, Chapter 7 constructs one experimental system which uses extreme points in scale space as feature points and SIFT to extract and match these points. The experiment demonstrates the feasibility and value of this method.(5)The fourth chapter addresses one instinctive and simple unified framework which unifies a lot of algorithms, such as SIFT, SURF and HARRI, and directs the reseach for extracting the invariant feature points.更多还原