多视点深度图像的配准方法研究与实现

【作者】 吕渭萍；

【导师】 张建中；

【作者基本信息】 电子科技大学 ， 计算机系统结构， 2008， 硕士

【摘要】 利用三维扫描技术来对物体三维重构是当今虚拟现实领域中的一项新技术,并且被广泛应用到案件现场三维重建、影视制作、文物保护等领域。为了重构三维物体的表面形状,必须事先得到物体表面的深度信息。但由于物体本身表面常常存在自遮挡和扫描角度误差等影响,只通过一次扫描不可能得到复杂物体表面的所有深度图像数据。必须通过扫描仪获得物体表面不同方向的深度图像;然后配准这些不同视角的深度图像,然后对其进行补洞、集成融合等操作,完成物体的三维建模。其中,深度图像配准是三维数字造型技术中的关键,它直接影响最后的合成结果、造型精度以及三维重构过程中的自动化程度。本文首先简单介绍了利用三维扫描仪获取的深度图像对物体三维重构的一般流程,然后介绍了深度图像对应点集配准的算法原理,提出一种改进的配准算法。ICP(Iterative Closest Point)算法是目前最常用的深度图像配准算法,但ICP算法对两幅待配准深度图像初始相对位置要求较高,它们之间的初始位置不能相差太大,否则,ICP的收敛方向是不确定的,因而配准结果也是不可靠的,ICP算法很可能收敛于局部最优点,故初始相对位置对结果具有决定性影响,为此人们提出了一些基于ICP的优化算法。本文通过重点分析ICP以及其变种算法的优缺点,提出了一种“由粗到精”的配准思路,将遗传算法与改进后的ICP算法相结合运用到深度图像配准过程之中。初始配准主要利用遗传算法强大的全局最优搜索能力以及问题域的独立性和应用的鲁棒性特征,采用实数编码、通过实验确定了遗传算法对于配准问题的参数设置,缩小两幅深度图像之间的位置差,达到提高ICP算法稳定性的目的;在接下来的二次配准过程当中,从重叠区域检测、控制点选取、对应点集计算和对应点有效性检查等方面分别对原始的ICP算法提出了多种改进措施,实现了两幅深度图像之间快速精确的配准。实验表明,该算法易于实现、对待配准图像初始位置以及重叠部分的大小没有严格要求,有很好的鲁棒性,并且配准过程完全自动化、不需人为手动干预,提高了三维重构的效率。本文最后在基于上述算法的基础上,实现了案件现场三维重建与智能分析系统中三维重建部分子模块。通过总结课题的研究工作,在本文最后还指明了下一步有必要深入开展的研究方向。更多还原

【Abstract】 Three-dimensional objects reconstruction with 3D-scanning technology is a new technology in the field of Virtual reality, and this technology is widely used in 3D reconstruction of crime scenes, Television production, Heritage preservation and other fields. To reconstruct the three-dimensional object, one must get range images from the surface of the object. Confined by eye direction and the shape of an object, range images can not be acquired to describe the object by one scanning. To acquire a complete surface model of an object, one must scan the object from different views firstly, register range images of different views then and finally merge all range images in an unified coordinate. The key in the above process is the registration of range images, which influences the final result, the precision and automatation of 3D reconstruction.The paper introduces the process of getting 3D models using range images obtained from the laser scanning simply, introduces the theory princeple of corresponding points registration and proposed an improved registration algorithm then. ICP(Iterative Closest Point) algorithem is the most widly used registration algorithem. Nevertheless, the proper convergence of ICP is guaranteed only if one of the datasets is a subset of the other; otherwise, erroneous alignments can result. Another drawback of ICP is that it requires a good pre-alignment of the views to converge to a correct solution. In order to solve these problems, researchers have proposed some improvements based on ICP. This paper focus on analyzing ICP algorithm, as well as the advantages and disadvantages of its variants,and proposed a "coarse-to-fine" registration thinking, combined Genetic Algorithm with improved ICP to the range registration. Rough registration mainly makes use of the Genetic Algorithm’s powerful global optimal search capabilities, the independence of the problem domain and the robust nature of its application. Using real-coded ,we can determine the GAs parameter settings based on experiments for range registration. After rough registration, narrowed the location between two range images, improved the stability of ICP algorithm that will be implemented. In the next registration process, the paper introduces a fast iterative pairwise registration method, which combines four acceleration techniques: fast detection of overlapped regions, more-careful selection of control points, fast surface closet point computation and compatibility test of pairing points based on the difference of modulation. Experiments show that the new algorithm is easy to implement, with quadratic precision, it is not strict for initial position and the size of overlapping part, and robust. The process of range registration is fully automated, without human intervention, and can increase the efficiency of 3D reconstruction. Finally, based on the above algorithms, implemented the 3D reconstruction module in reconstruction and intelligence analysis system of crime scenes.After summing up the research work, this paper points out some future research works.更多还原