基于双目立体视觉的三维场景重建研究

Research on Binocular Stereo Vision Based 3D Reconstruction of Scene

作者: 专业:信号与信息处理 导师:胡家升 年度:2010 学位:硕士  院校: 大连理工大学

Keywords

Stereo Vision, Stereo matching, Graph-cut, Minimization of Energy Function, Min-cut, Mean Shift

        基于双目立体视觉三维重建系统是通过双目摄像机在同一场景下取得不同视角的两幅图像来恢复场景三维信息的过程。此种方法是被动三维测量的一种重要方法,在许多领域都有着广泛的应用,如:遥感成像,机器人导航等。基于双目立体视觉三维重建的终极目标是通过图像平面信息来恢复出物体的三维模型,其中物体表面点深度的获取就成为整个过程的最为核心部分,而物体表面深度信息则是通过图像的立体匹配获取的。现有的立体匹配方法,可分为两类:基于窗口的局部算法和基于能量函数最小化的全局算法。由于全局立体匹配算法所得结果的精度要远远的高于局部算法,因此受到了国内外学者的广泛关注。在基于能量函数最小化的全局算法中,基于图割(Graph-cut)和基于置信传递(Belief-propagation)两类能量函数最小化算法在应用上表现的十分突出,精确性得到了广大学者的一致认可。然而,在基于立体视觉三维重建系统中,要想通过图像来获取高分辨率的三维模型仍然是十分困难的。上述研究背景下,本文对基于双目立体视觉三维重建系统进行论述,并对立体匹配和场景重建两个重点环节做了深入的研究,目的是提高现有算法的精确性和鲁棒性,研究的具体内容有以下几个方面:在立体匹配方面,本文采用的是基于图割的立体匹配方法,这种方法首先根据图像对之间的彩色特征差异和相邻区域间应该满足的视差平滑约束来定义立体匹配能量函数。根据能量函数的形式来构造与能量函数相对应的图,然后对所构造的图求取最小割运算。求得图的最小割即实现了能量函数的最小化,此时对应的视差分配结果即是立体匹配的最佳结果。针对基于图割的立体匹配算法得到的视差图中存在大量匹配噪声和误匹配视差点的问题,本文引入了均值平移算法。该方法联合了图像的彩色信息和视差信息,以图割立体匹配算法得到的视差图作为初始视差图,对初始视差图进行优化,消除了视差图中的匹配噪声和误匹配视差点。同时,由于均值平移算法图像分割具有很好的边界保留效果,场景中物体的边界得到了很好的优化,经过局部场景视差图的边缘检测可以清楚的看出本文的改进效果。在场景的重建方面,本文对优化后的立体匹配结果进行三维点的计算,给出了恢复的三维场景。最后,对三维场景进行贴纹理操作,将数字图像以像素为单位映射到对应的三维场景点上,大大提高了三维场景的真实感。
    Stereo vision is recovering 3d depth information of objects in a scene using image pairs taken from two different views through one camera or two. This method is one of the most important passive methods in stereo vision. There are many applications in different areas, such as Stereo-imaging in Remote sensing, Robot navigation and etc. The final goal of stereo vision is to get 3d model from image pair. So getting the depth of the object surface is the key step in the whole process, the depth information is getting through stereo matching. In general, stereo matching can be classified into the local or global approaches. The result of the global approaches achieves high accuracy, and many researchers pay attention to the global approaches. In global approaches, Graph-cut based algorithm and Belief-propagation based algorithm are prominent in applications. Their accuracy is accepted by the researchers. However, due to the ill-posed nature of stereo matching problem, determination of accurate is still a hard problem.In the paper, we lucubrate two aspects of stereo vision system, that are stereo matching and scene reconstruction, the main research content are as follows:In stereo matching, using an Graph-cut based algorithm, this algorithm construct energy function according to color differences and smoothness constraint between neighbor disparity areas, then construct graph corresponding to the energy function and get the min-cut on graph. That is getting the min-value of the energy function and we get the best disparity map.There are many noises and local error disparity pixels in the disparity map got from basic Graph-cut based stereo matching algorithm above. In order to improve the accuracy of disparity map, we introduce mean-shift filtering technology to the basic stereo matching method, our method considers the segmentation using both color and disparity information simultaneously to improve the initial map achieved by the basic algorithm. In the result, our method is so effective in removing the noises and error pixels improve and improving the accuracy of the boundaries in scene, through using edge detection on local disparity map. We can easily find the effective of the algorithm.In scene reconstruction, we compute the depths of the object after stereo matching and get the 3d model of the scene, through texture wrapping we paste the reference image to the 3d model and improve the reality of the 3d scene.
        

基于双目立体视觉的三维场景重建研究

摘要4-5
Abstract5
1 绪论9-19
    1.1 立体视觉系统9-10
    1.2 立体视觉三维重构研究历史和现状10-13
        1.2.1 摄像机标定方法的发展11-12
        1.2.2 图像立体匹配算法的发展12
        1.2.3 三维重构方法发展12-13
    1.3 三维重构的主要方法分类13-17
        1.3.1 接触式测量法14
        1.3.2 结构光法14-16
        1.3.3 双目立体视觉图像重建方法16
        1.3.4 其他三维重建方法16-17
    1.4 论文的主要工作内容和组织形式17-19
2 三维重构系统的基本理论19-26
    2.1 摄像机标定19-21
    2.2 成像模式和视差场21-23
    2.3 立体匹配中的基本假设及约束条件23-26
        2.3.1 立体匹配中的基本假设23-24
        2.3.2 立体匹配中的约束条件24-26
3 基于图割的立体匹配算法26-39
    3.1 通过像素标号来解决立体匹配问题26-27
    3.2 能量函数27-29
        3.2.1 能量函数的结构27-29
        3.2.2 二进制能量函数29
    3.3 扩展移动算法29-31
    3.4 图的构造31-35
        3.4.1 图基础知识31-32
        3.4.2 根据能量函数构图原则32-33
        3.4.3 构图过程33-35
    3.5 最小割/最大流算法35-39
        3.5.1 算法概述35-37
        3.5.2 算法流程37-39
4 基于图割和均值平移的立体匹配新算法39-45
    4.1 均值平移算法基本原理39-42
        4.1.1 均值平移向量39-41
        4.1.2 均值平移算法41-42
    4.2 均值平移算法图像分割42-43
    4.3 立体匹配中的均值平移算法43-45
5 实验结果与分析45-56
    5.1 三维形貌恢复45-46
        5.1.1 视差图与深度45-46
        5.1.2 纹理映射46
    5.2 立体匹配实验及分析46-51
    5.3 三维重建实验51-56
总结与展望56-57
参考文献57-59
攻读硕士学位期间发表学术论文情况59-60
致谢60-62
        下载全文需10


本文地址:

上一篇:基于鼾声检测的睡眠呼吸暂停低通气综合症诊断
下一篇:基于像素级空间金字塔和乘法融合的目标跟踪

分享到: 分享基于双目立体视觉的三维场景重建研究到腾讯微博           收藏
评论排行
公告