【作者】 高鹏东；

【导师】 彭翔；

【作者基本信息】 天津大学 ， 测试计量技术及仪器， 2007， 博士

【摘要】 近十年来,随着计算机硬件技术的不断进步,国际上对三维数字成像及造型（3D Digital Imaging and Modeling, 3DIM）技术的研究得到了飞速发展,该技术的应用也得到了更广泛的普及。作为一门涉及光学、计算机技术以及微电子技术等多学科交叉的新兴领域,3DIM技术的快速发展,正在从多个角度改变着人类的生活。从传统的工业制造到文教事业中的文物保护,可以看到各种各样3DIM技术的应用实例。而深度像配准和曲面细分是3DIM技术的两个核心内容,因此对这两种技术以及3DIM技术应用的研究具有重要的理论和现实意义。深度像配准是三维数字成像的关键一环,其目的就是寻找不同视角深度像间的空间位置转换关系,从而将这些深度像统一到同一个世界坐标系内。然而,现在应用最为广泛的深度像精确配准方法——ICP算法却始终受到两个关键问题的困扰:对初值依赖和对噪声敏感。因此,本文首次提出了基于表面间平均体积测度的概念,并基于该测度实现了ICP和遗传算法两种不同框架下的迭代配准。实验结果表明,本文提出的这种测度比现有基于点对对应的测度更加鲁棒,更能反映深度像重叠区域的吻合程度,抗噪声能力也更强。在进行三维散乱数据的曲面拟合时,面对曲面光滑性、准确性以及多分辨的表达等要求,细分曲面成为一个有效的解决方法。但是如何保持曲面上的各种尖锐特征是细分技术应用的关键。现有的拟合系统,使用的都是Hoppe改进的Loop细分模式。作为一种采用1-4分裂的细分方法,Loop模式所能提供的多分辨层次非常有限。因此,本文改进了现有的3^2/1细分模式,提出了一种分片光滑曲面的重建方法,并给出了改进后3^2/1细分极限曲面的连续性分析。结果表明该方法具有较高的曲面拟合精度和较强的多分辨表达能力。3DIM技术的飞速发展,为真实感三维人脸建模提供了更多的新技术以及更广泛的应用。因此,本文结合曲面造型技术的最新发展,提出了一种基于两张正交照片和一个属性模型的真实感三维人脸建模方法。把自由曲面变形、网格简化以及曲面细分等技术综合起来,构造了一种新的几何建模方法。同时,为了提供个性化的人脸信息,还提出了一种基于模糊模式匹配的人脸检测和定位算法,以及一种基于改进的彩色Snake模型和形态学边缘检测的轮廓线提取算法,从而保证了真实感三维人脸建模的有效实施。更多还原

【Abstract】 During the past decade, as the progress of computer hardware technology, the research on three dimensional imaging and modeling （3DIM） has been increasingly active while the applications of 3DIM have also appeared popular ever since. This tendency attributes to the fast development in the field of optical engineering, computer science and technology, as well as the microelectronics technology. Nowadays, 3DIM has gradually changed our daily life ranging from traditional manufacturing industry to preservation of historical heritages. The contents of this thesis addresses two key issues involved in three dimensional imaging and modeling, namely, range image registration and surface subdivision. The research on them as well as on the application of 3DIM is meaningful in both theory and practice.Range image registration is one of the key issues in three dimensional imaging and modeling. It aims to find out the rigid transformations of all range images captured from different viewpoints and then bring them into a common coordinate system, i.e. the world coordinate system. However, the most popular registration method, ICP algorithm, is always suffered from two serious problems: initial pose estimation and noise outliers. Therefore, we present a new concept, Surface Mean Inter-Space Measure, in this thesis for the first time in order to assist accurate registration. On the basis of the new measure we have fulfilled the registrations under two different frames, ICP and GA, respectively. The experimental results illustrate that the proposed measure is more robust than the point-corresponding method. And it’s more insensitive to noises as well as more precise in measuring the quality of registration results.The second key issue to be addressed is the subdivision surface, which is an effective tool for fitting the three dimensional scattered data. Subdivision surface can be used to solve the problems such as the smoothness, precision, and level of detail modeling as well. However, there is still a constraint in conventional subdivision techniques. For example, it is hard for them to preserve the sharp features in original models. The scheme now used in most existing surface fitting systems is the Loop method modified by Hoppe. With a 1-4 split operator, this scheme would achieve limited level of details. In order to overcome this limitation, we develop a new approach that improve the rules of 3^2/1 subdivision and reconstruct the piecewise smooth surfaces. The continuity of modified subdivision surface is also analyzed. The results indicate that the presented method can introduce more levels of details and provide a better quality of fitting surfaces.The third part of this thesis involves realistic 3D facial modeling. The fast development of 3DIM brings more techniques and applications for realistic 3D facial modeling. So in this part, we present a new method for this purpose. The proposed method makes use of two orthogonal photographs and a generic facial model to reconstruct photo-realistic 3D facial models. The free-form deformation and mesh simplification are incorporated with the subdivision techniques to construct a new method to build the geometric models. At the same time, a new face detection algorithm based on fuzzy pattern matching and a new facial feature contour extraction method based on a modified color Snake model and some morphological edge detectors are presented to provide individual facial information. Both of them are necessary for realistic 3D facial modeling.更多还原