CASIA OpenIR  > 毕业生  > 硕士学位论文
基于双目视觉的背景分割技术研究
其他题名Background Segmentation with Binocular Vision
时岭
2010-05-31
学位类型工学硕士
中文摘要背景分割技术是计算机视觉领域的一个重要研究方向。背景分割在数字娱乐,影视制作和人机交互等领域具有广泛的应用前景。三维重建技术是计算机视觉的一个基本目的,主要方法是通过两个以上的图片的匹配和运算,来理解场景或物体的三维结构。本文在对三维重建技术研究和实验的基础上,结合场景的双目深度信息以及基于背景建模的动态图切技术,以视频中的前景和背景的分割技术为研究主体,对背景分割的原理的方法进行了实验和分析。同时,在分割的基础上,完成了二维平面上基于的人体躯干的运动捕捉。本文的主要贡献如下: ① 实现了基于图像分割和置信传播技术的立体匹配方法,提出了基于熵值和双层分割的多角度立体匹配方法,从而得到更加准确和平滑的深度图。在双目视觉的密集匹配中,本文采用了改进的循环信息传播的方式,在较小的时间代价下得到准确的深度信息。在多角度的三维重建过程中,利用已知的摄像机内外参数,在重建对象的位置和角度都有较大差异的时候,获取到比较准确和平滑的深度图。并且,在两次分割(粗分割和过分割)进行平面拟合的基础上,有效的去除了噪声,得到优化的深度图。 ② 实现了基于背景建模的单目视觉背景分割。讨论了基于双目视觉进行背景分割的研究现状和发展前景,提出并实现了一种将双目的信息融入到单目图切的方法。该方法利用双目视觉获得的立体信息,改进了图切算法中的初值,使得图切算法可以利用深度信息剔除容易受到干扰的区域。在此基础上,通过边界的跟踪和平滑进行了后处理,取得了获得更好的分割效果。 ③ 提出了一个基于双目视觉的人体运动捕捉系统。在背景分割的基础上,假设场景中的前景是人体躯干,将分割出来的前景进行处理,并且得到一个人体躯干的骨架,从而完成在二维平面上的运动捕捉。
英文摘要Background/Foreground segmentation is one of the most important fields in computer vision, which has a wide range of applications including digital entertainment, film/TV industry and human-computer interaction, etc. 3D reconstruction is one of the main goals of computer vision. With two or more images from different perspective points and appropriate algorithms and parameters, computer can understand the 3D information of the special scene or object by 3D reconstruction algorithms. In this paper, we focus on the combination of 3D reconstruction and the algorithm of Graph Cut based on background modeling, which makes the advantages of 3D information provided by binocular vision and segmentation ability of Graph Cut by Max-flow/Min-cut algorithm. With background segmentation results, we made the motion capture of human body as foreground. Results of these methods are shown in this paper. The main contributions of this paper are as follows: ① We implemented the stereo matching algorithm based on color segmentation and loopy belief propagation, and proposed a multi-view reconstruction method based and entropy evaluation and two-stage segmentation. In binocular stereo vision, we used loopy belief propagation for dense matching and got accurate depth information. In multi-view stereo, with know intrinsic and extrinsic parameters of the camera, we achieve accurate and smooth depth maps even when the scene has big position and orientation differences. Besides, with two-stage segmentation (coarse/over segmentation), we made plan fitting of the depth maps for optimized results. ② We implement background segmentation with monocular vision and discussed the research of segmentation with binocular vision. Based on this, we proposed a combined algorithm for improved Graph Cut, which add the stereo information into the initialization of Graph Cut. With the hypothesis that background has bigger disparity than the foreground in binocular vision, we made use of stereo information which is superior to the color information and background in some parts of the scene. Besides, we used some algorithms for better segmentation results with boundary tracking and boundary smoothing. ③ With the segmentation results, we implemented a motion capture system of human body under the hypothesis that the foreground only consists of human body. We simplified the foreground silhouette and got the key points of human skeleton and get the basic motion capture of human body in 2d image.
关键词三维重建 立体匹配 背景分割 背景建模 运动捕捉 3d Reconstruction Stereo Vision Background Segmentation Background Modeling Motion Capture
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/7546
专题毕业生_硕士学位论文
推荐引用方式
GB/T 7714
时岭. 基于双目视觉的背景分割技术研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2010.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
CASIA_20072801462802(2759KB) 限制开放CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[时岭]的文章
百度学术
百度学术中相似的文章
[时岭]的文章
必应学术
必应学术中相似的文章
[时岭]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。