【作者】 谢丰；

【导师】 张鸿宾；

【作者基本信息】 北京工业大学 ， 计算机应用， 2003， 硕士

【摘要】 获取真实物体的三维模型是计算机视觉的一个重要研究课题，在虚拟现实、CAD反向工程、模式识别、不良产品检测以及人体非接触测量等领域有着广泛的应用。本文采用相位法从多个不同的视点获得距离图像数据，通过多视点距离图像的对准获得视点之间的运动参数，然后经过距离图像的集成，建立完整的物体表面的几何形状描述。 本文首先介绍了有关三维距离传感技术的原理和我们采集距离图像的环境，并且简单分析了采集的距离数据中包含的误差。然后提出了一种基于三角形匹配的多视点距离图像的对准算法。在现有的代表性的对准算法中，都是把距离图像看成3D点集，通过匹配不同距离图像之间的点对，估计运动参数。其主要的缺点是没有充分利用距离图像上象素点之间的邻接信息。本文算法针对这一缺点采用三角形网格最近距离的平均值为评价函数。通过在三角形网格上按面积均匀分布的假想抽样，在没有明显增加计算复杂度的情况下，使得代入ICP算法的3D点集，不仅是近似均匀地分布在物体表面上，而且更加充分地反映了被测物体的表面信息。实验表明，该对准算法具有对初始运动参数的要求不高，收敛速度较快，较高对准精度和较强抗噪声能力的优点。 本文的集成算法是对zippered算法的改进，它对删除冗余三角形后的原始三角形网格做连通性的检查，通过删除那些三角形个数小于一定阈值的子连通网格，来达到在两原始三角形网格缝隙附近去除那些相对于被测物体表面有较大偏差的三角形的目的。在充分考虑删除冗余三角形后的两原始三角形网格缝隙之间的各种情况，本文对各种情况有针对性地提出了具体的处理办法。实验表明，本文的集成算法在计算时间和生成表面的质量上都可以满足三维模型重建的需要。 本文提出的对准和集成方法为自动几何建模提供了一种较为有效的途径。更多还原

【Abstract】 Reconstruction of 3-D objects is an important research problem of computer vision, which have been applied in many domains such as Virtual Reality, CAD reverse engineering, Pattern Recognition, ill product detection, and non-contact measurement of 3D human body. In the paper, the problem of registration and integration of multiple range images is addressed. Multiple range images can be obtained by turning the 3D object at different angles around sensor. The range images of different viewpoints are registered to estimate the relative transformations among the viewpoints, and then integrated to obtain a complete 3D geometric description of object.First, we introduce the principal of three-dimensional sensors and our data collecting system, and analyze two kinds of data errors in the range images. Next we propose a novel matching triangles registration algorithm. In some representative algorithms for the registration of multiple range images, the range image is only considered as a 3D point set. The motion parameters are estimated by matching the closest point in other range image. Their main drawback is the registration process doesn’t make full use of the information of range images. So our algorithm defines cost function is the mean distance of two triangle meshes. By an imaginary uniform sampling on the original triangle mesh, our algorithm makes the 3D point set using in ICP algorithm not only approximatively uniformly distribute in the measured object surface, but also possess of more information of the object surface than classical ICP algorithm. The experimental results show the proposed registration method is computationally efficient and robust to outliers and initial motion parameters.Last we propose an algorithm for integration of multiple range images, which improves the zippered algorithm by inspecting the connectivity of the original triangle mesh after the procedure of removing redundant surfaces. Moreover we carefully consider all possibility of the gap of two original triangle meshes after the procedure of removing redundant surfaces, and then propose corresponding methods to deal with all these possibility. Experiments demonstrate that the method can satisfy the requirements of computing cost and accuracy of object model reconstruction.Our registration and integration methods provide an easy way to fastautomatically digitize 3D object model.Xie Feng (Pattern Recognition and Image Processing)supervised by Prof. Zhang Hongbin.更多还原