【作者】 刘晓利；

【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】 三维数字成像及造型(3DIM)作为一个新兴领域,覆盖了光学/光电子、精密测量、计算机视觉以及计算机图形学等多学科知识,其关键技术有:三维光学传感、系统标定、深度像匹配、深度像融合、三维模型简化、细分、参数化及三维模型编辑等。本论文主要围绕多视场深度像的处理上的若干关键技术进行探讨,主要有深度像的匹配、融合及简化技术。深度像匹配是多视场深度像造型的关键一环,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坐标系内。深度像匹配可大致分为:两个视场的深度像粗匹配、精匹配和多视场的深度像全局匹配。本文提出了结合纹理信息进行深度像匹配的方法,由于二维纹理匹配技术较成熟且不受几何噪声影响,可以自动、快速完成深度像的粗匹配,实验结果表明所提出的算法具有一定的鲁棒性。目前应用最为广泛的深度像精匹配方法是ICP算法,该算法依赖一个初始估计、对噪声敏感且计算复杂度高。因此,本文提出并研究了非编码标志点进行多视场深度像的匹配方法,并结合全局优化算法避免了匹配误差积累,获得了较好的实验结果。此外,本文还详细比较了该方法与ICP方法的差异及改进方法。深度像的融合是多视场深度像造型的另一重要环节。目前被广泛应用的深度像融合算法主要有基于网格缝合和基于隐式曲面的两类方法,其各有优缺点。一种快速、鲁棒、高精度的深度像融合算法还有待进一步研究。本文提出基于射线投影的深度像融合算法,利用轴向包围盒树结构求取交点,并借助Dexel结构进行存储,既加快了计算速度又节省了存储空间。实验验证了该算法的有效性。曲面简化技术在三维造型领域占有重要地位。本文概述了目前各种简化算法的各自特点及适用范围。并着重对二次误差测度(QEM)简化算法进行改进,提出顶点尖特征度的概念,在测度函数上加入一个惩罚项,从而改变边折叠的次序,使得原始模型的尖锐特征在最终简化结果中得以保留,并且网格能够合理分配,在曲率较低区域分布稀疏,曲率较高区域分布密集。更多还原

【Abstract】 As a new realm, the content behind the Three-dimensional Imaging and Modeling (3DIM) involves multi-disciplines such as optics and optoelectronics, precision metrology, computer vision, computer graphics and so on. In hiberarchy, there are several key components regarding to the 3DIM, which include 3D optical sensing, 3D system calibration, range image registration and integration, 3D modeling and so forth. This dissertation focuses on the three key issues with respect to 3D modeling of multiple range images, namely, the range image registration, integration and the 3D model simplification.Range images registration is of a critical issue in 3D modeling. The registration of the range images taken from different view points aims to find the best estimate of rigid transformations among the pairwise range images and then put them into a common coordinate system. In terms of the registration the existed approaches can be classified into three distinct levels: the coarse registration, the fine registration of pairwise range images and the global registration of multiple range images. We propose a novel method for the coarse registration with the help of texture information by observing the fact that the 2D texture registration is not affected by geometry noise and can be performed rapidly. The experimental results have verified the efficiency and robustness of proposed technique.Furthermore, the most popular fine registration method is on the basis of the iterative closest point (ICP) algorithm, which was subjected to the four limitations: initial estimate for rigid transformation was needed, fragile to the noise or the outliers, error accumulation and with great time complexity. To overcome these disadvantages, in this thesis, we present a new fine registration method to register the multiple range-images with non-coding markers. In the proposed algorithm, a global optimization technique has been employed to reduce the error accumulation and the sensitivity to the noise and outliers. The experimental results illustrate that the proposed method is very efficient and time-saving compared with previous methods. In addition, we also make a thorough analysis on the difference between marker-based algorithm and ICP-based algorithm.Integration or data fusion of multiple range images is another paramount technique of 3D modeling. Up to now most of popular integration methods have been based on mesh stitching or implicit surface. All of these methods had their pros and cons either time-consumig or difficult to deal with real object surfaces with complex geometry and topology, and these methods were designed specifically for different applications. However, a fast, robust and high-accurate integration method need to be further developed. This thesis explores a new algorithm based on ray-casting, in which an axis-aligned bounding box tree is used to compute the intersections, and Dexel data structure is employed to save storage space. The experiment results show that a good range data fusion results can be obtained with the proposed method.Model simplification also plays a key role for 3D modeling, which would be viewed as an equivalent operation to the data compression in its counterpart of image processing. This thesis gives a comprehensive review to the most notably simplification algorithms already available in this field. Then we introduce a concept of sharp degree of vertex. With the concept of sharp degree of vertex, we can make an improvement to the quadric error metric (QEM) algorithm. In our approach, an penalty term is added to the measure function to make possible for adjusting the order of edge collapse. With proposed method, we are able to maintain the sharp feature of original model while achieving a large compression ratio in the simplified model. Experiment results show that the proposed method can give a better simplified model with dense meshes in the high-curvature regions and sparse meshes in the flat regions.更多还原