基于多视图三维重建技术研究
文献类型：学位论文
| 作者 | 黄以君 |
| 学位类别 | 博士 |
| 答辩日期 | 2010-05-29 |
| 授予单位 | 中国科学院沈阳自动化研究所 |
| 授予地点 | 中国科学院沈阳自动化研究所 |
| 导师 | 刘伟军 |
| 关键词 | 计算机视觉 参数估计 多视图 三维重构 三维测量 基于图像的三维建模 |
| 其他题名 | Research on 3D Reconstruction Based on Multi-Images |
| 学位专业 | 机械电子工程 |
| 中文摘要 | 随着现代计算机技术的飞速进步，计算机图形、图像处理、辅助设计多媒体技术越来越广泛深入地应用于工业、国防、医学、影视业、广告等各个领域。人们经常需要能迅速地获得物体表面的三维信息，将其转变成计算机能直接处理的数据，基于计算机视觉的三维信息获取技术具有不可替代的重要作用。 利用多视图进行三维重建是计算机视觉的热点问题，它融合了多幅图像中的信息，在三维物体测量和创建高度真实感的三维模型方面具有巨大的优势，在诸多领域中具有极其广阔的应用前景。 但是，基于多视图的三维重建技术超越了传统视觉的研究范畴，需要建立新的理论并采用新的解决方法，因此这种技术还不够成熟。本文旨在探索利用多视图实现三维重建的新方法，深入研究一些关键问题，以推广该技术在实际中的应用。主要内容包括： 研究计算机视觉中的参数估计问题。从线性、非线性和鲁棒性三个方面进行讨论，为参数估计提供统一的分析框架；分析了视觉中常用的鲁棒估计方法的基本思想和局限性，引入了统计回归分析中的LQS和LTS估计方法，克服了测量误差方差的先验知识对计算结果的影响，并且对测量数据的准确度没有要求，适用于野值比例大的情况；采用Bucket分割技术，避免了随机抽样检验时样本局部集中的问题；以计算基本矩阵为应用实例，使用本文提出的鲁棒估计方法解决具体参数问题。与常用的估计方法相比，本文方法具有更高的计算精度和鲁棒性。 研究利用多视图几何关系的度量重建方法。分析了多视图之间的对极几何关系，将传统的两视图的度量重建问题推广到多视图；将多线性关系用于场景的三维重建中，通过对重建过程进行详细推导，提出了统一实现框架；使用直接重建方法，利用多焦点张量恢复相机的相对位置，得到场景的度量重建；分析并比较了使用二、三、四焦点张量重建方法的计算精度和效率。基于三焦点张量的重建方法可以应用于三维精确测量中，具有很大的工程应用价值。 研究基于图像的三维建模方法。提出了一种由图像中的二维轮廓构造物体空间包围壳的快速方法，结合具体问题、利用坐标变换，将复杂的三维求交转化为一系列简单的二维操作，通过面片拼接得到物体的空间包围壳，作为进一步优化建模的初始模型；分析了Space Carving建模理论的特点和缺陷，采用联合阈值模型，避免了Space Carving方法中单一全局阈值的约束，并使用三组正交平面表示体素空间，便于从多个方向扫描三维空间，实现全方位重建；分析了判断体素一致过程中的统计特性，提出了一种基于概率的一致判断函数，具有更高的鲁棒性；利用本文提出的建模方法，能够由图像序列重建出物体的完整且具有照片真实感的三维模型。 |
| 索取号 | TP391.4/H79/2010 |
| 英文摘要 | Recently, with the rapid development of the computer science, the computer graphics and images processing technology, the aided design technology and the multimedia technology have been applied to industry, national defense, medicine, film and TV industry and advertisement widely. Frequently the 3D information on the object surface need to be obtained rapidly and transformed to the data which can be disposed by the computer. The technology of 3D information acquisition based on the computer vision plays an irreplaceable role. Reconstructing the digital model of an object from multi-views has been a hot topic in the field of computer vision. By fusing the information in multi-images, this technology has enormous advantages in 3D accurate measurement and realistic modeling, therefore, has a wide application field and good prospect. But the technology for 3D reconstruction of multi-images is beyond the research category of the traditional computer vision and needs novel theories and resolvent methods, so this technology is not mature. This thesis studies and discusses the new approaches and some critical problems in 3D reconstruction based on multi-images in order to extend its application in the practical engineering. Major contributions are as follows: The parameter estimation in the computer vision is researched. The paper discusses from the linear, nonlinear and robust discussion of three aspects and provides a unified analytical framework for the parameter estimation; analyzes the basic ideas and limitations of common robust methods in computer vision and introduces LQS/LTS regression to overcome the shortcomings of the traditional methods that the prior knowledge of the measurement error variance effects the calculation results in the and the outliers in the initial data are not more than half of the entire data set; adopts the bucketing technique during the process of sampling for regularly random selection avoiding the uneven distribution of sample data; estimates fundamental matrix robustly from matching points to exemplify how to estimate parameters using the proposed method. Experiments for both synthetic and real images show the accuracy and robustness of our method. The problem of reconstructing metric 3D scenes is investigated from the point of view of the multi-view geometry. By investigating the epipolar geometry between multiple views, the paper applies the multi-linear relations to the reconstruction of 3D scenes and develops a uniform reconstructing framework which is derived in detail. The framework carries on reconstruction directly, utilizes the tensor to analyze the relative positions of views, and obtains the final metric reconstruction of the 3D scene. Experimental results give a comparison of the precision and efficiency of the reconstruction algorithms respectively using two, three and four focal tensors and shows that the proposed method based on three-focal tensor could be used for accurate 3D measurement and has great engineer value. The method of image based modeling is discussed. The spatial bounding hull as the initial volume is quickly constructed from the contours in the images by a smart strategy which utilizes the coordinate system transformation to translate the volume intersection to a series of 2D operations and joints the facets into a close the spatial bounding hull. The paper improves the multi-threshold combined model, which overcomes the limitation of using a single global threshold parameter of the original Space Carving theory; adopts three sets of orthogonal planes to represent the voxel space, which is effective to multi-pass carving and reconstruct a complete model; analyzes the statistical characteristics of photo-consistency of the voxel and proposes a probability consistency function having the steady performance at different resolutions. The proposed method could obtain the complete and realistic digital model of a geometrically complex object using a set of easily obtained photographs. |
| 语种 | 中文 |
| 公开日期 | 2012-07-27 |
| 产权排序 | 1 |
| 分类号 | TP391.4 |
| 源URL | [http://ir.sia.ac.cn/handle/173321/9298]  |
| 专题 | 沈阳自动化研究所_装备制造技术研究室 |
| 推荐引用方式 GB/T 7714 | 黄以君. 基于多视图三维重建技术研究[D]. 中国科学院沈阳自动化研究所. 中国科学院沈阳自动化研究所. 2010. |
入库方式： OAI收割
来源：沈阳自动化研究所
浏览0
下载0
收藏0
其他版本
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。