【作者】 杜鹏；

【导师】 华蓓； 叶豪盛；

【作者基本信息】 中国科学技术大学 ， 计算机软件与理论， 2011， 博士

【摘要】 随着三维扫描和捕捉设备以及计算机建模工具的发展,可变形三维物体数据变得容易获取,并在广泛的领域得到应用。怎样对这些可变形三维物体进行统计建模以辅助应用已成为一个研究热点。本文提供有效的方法对两种不同形态(包括三维表面形态和三维骨骼形态)的可变形三维物体进行三维统计可变形模型(Statistical Deformable Model, SDM)的构建,并分别用于三维物体重建和三维运动(或动作)识别。对于表面形态的三维物体,由于其数据维度高,在建立SDM时往往会碰到小样本问题。为了解决该问题,本文基于分治的思想构建分块的SDM(PSDM),以替代全局单一的SDM。为了构建一个PSDM,其中有两个关键的步骤：(1)将物体表面划分为多个局部小片块；(2)将各个变形后的局部SDM装配成最终的SDM,以形成完整的物体表面。对于两种不同类型的三维表面数据,本文采用两种不同的技术分别为它们建立PSDM。一方面,来源于CT图像的三维医学表面数据有其特殊的分层结构,本文为其构建一个层次化的PSDM,该模型由一个建立在物体表面特征点上的粗略的全局SDM和一组建立在局部片块上的局部SDM组成,其中,全局SDM捕捉物体的全局变形特征,并提供一个框架以辅助物体表面的划分及局部SDM的装配,各局部SDM捕捉物体的局部变形细节。另一方面,一般三维物体表面数据没有特殊的结构,相对较难处理,对于其表面划分问题,本文基于物体表面可变性特征的相似度来进行划分,并提出两种新的度量标准来量化可变性相似度；对于其装配问题,本文采用一种基于约束变形的技术来对变形后的各局部SDM进行无缝粘接。为以上两种类型的三维物体表面数据构建的PSDM均应用于三维物体重建。另外,为了确保整个PSDM的全局形状一致性,本文进一步采用基于多级SDM的技术用于约束这些局部SDM的变形。对于骨骼形态的三维物体,如三维动作捕捉数据,本文为每种类型的运动构建一个行为(或类型)特异SDM来提取其共同特征,该行为特异SDM能够捕捉并描述每种类型运动的所有容许变形信息,利用此特性,本文提出根据各个行为特异SDM表达某个新运动的好坏程度(即各个行为特异SDM重建该运动的正确度或重建精度)来进行该运动的分类。本文展示该新方法比传统的基于本征运动的方法更适合于三维运动分类。此外,本文还描述一种新的基于统计变形特征的三维运动相似度度量方法,并将该方法用于三维运动分类。更多还原

【Abstract】 With the development of3D scanner and capture devices and computer modeling tools, deformable3D objects have become easy to obtain and have been used in a wide spectrum of fields. How to statistically modeling these deformable3D objects for various applications has become an active research topic.This thesis proposes sophisticated methods to construct3D Statistical Deformable Models (SDMs) for two different forms (i.e.,3D surface form and3D skeleton form) of deformable3D objects for object reconstruction and motion recognition, respectively.For the3D surface form objects, the small sample size problem is frequently encountered when constructing SDM for them, due to their high data dimensions. To address this problem, this thesis constructs piecewise SDM (PSDM) based on divide-and-conquer strategy instead of single global SDM for the objects. To construct a PSDM, two key steps are required:(1) partitioning the surface into multiple components, and (2) assembling the deformed local SDMs to form the final SDM for the object surface. Studying on two different kinds of3D surface data, we propose different techniques for constructing PSDMs for them, respectively. On one hand, the3D medical surface data derived from CT images has a special multi-layer structure that is relatively easier to process. We construct a hierarchical PSDM for it, which consists of a coarse global SDM built on feature points of the surface and a set of local SDMs built on local surface components. The global SDM serves to capture the global variability of the object, and provide a framework for partitioning the surface and also for assembling the local SDMs. The local SDMs serve to capture the local deformation details. On the other hand, the generic3D surface data has no specific structure and is much more difficult to deal with. For the surface partitioning issue, we partition a surface based on the similarity of the surface variability characteristics, and subsequently propose two novel measures for quantifying the variability similarity. For the assembly problem, we employ a technique based on constrained deformation for seamlessly stitching the deformed local SDMs. The PSDMs for the two kinds of3D surface data are both applied to3D object reconstruction. For ensuring the global shape consistency of the entire PSDM, we further propose a multi-level SDM based technique to constrain the deformation of the local SDMs. For the3D skeleton form objects, e.g., the motion capture data, we construct a behavior-specific SDM for each type of the motions in order to capture the common characteristics shared by the motions of that type. The behavior-specific SDM is able to capture and encode all allowable deformation for the type of motions it represents. Taking this property, we propose to classify a new motion based on how well each behavior-specific SDM represents it, i.e., how accurately each behavior-specific SDM can reconstruct the new motion. We show this novel technique is more powerful for3D motion classification compared with the traditional eigen-motions based classification technique. In addition, we also present a novel statistical variation characteristic based measure for quantifying the similarity of3D motions, and apply it to3D motion classification.更多还原