【作者】 韦虎；

【导师】 张丽艳；

【作者基本信息】 南京航空航天大学 ， 航空宇航制造工程， 2010， 博士

【摘要】 对现实世界的物理模型进行三维外形数据采集并构建其数字化模型,在航空、航天、汽车、船舶、机械制造、生物医学、游戏娱乐等行业有着广泛的应用需求,而对测量数据的处理是架设在数据采集与数据应用之间的桥梁,是三维外形测量技术中的重要环节。随着数字图像处理技术快速发展,以面阵摄像机为主要传感器件的三维外形测量技术可以快速获取物体表面的深度图,产生的稠密点云能在较高的分辨率下刻画被测物体上的细节特征。特别是近年来芯片制造技术的迅速发展、相应产品的价格不断下降,使得基于面阵摄像机的三维外形测量技术得到了长足的发展。本文深入研究了基于面阵摄像机三维外形测量系统中数据处理的若干关键技术,包括数据预处理、深度图粗配准、深度图精配准、深度图融合以及网格模型的光顺与简化。本文的主要内容和创新点总结如下:1、针对基于面阵摄像机的测量系统所获得的海量数据同时具有深度信息和像素结构信息的特点,分别提出了稠密点云三角化和自适应取样三角化方法。其中稠密点云三角化方法可以快速得到测量数据的全分辨率网格模型;自适应取样三角化方法通过对测量数据的不同区域进行不同比例取样,在简化显示测量数据的同时保持物体三维外形轮廓,解决了海量测量数据的实时显示问题。2、在对已有粗配准算法进行深入分析的基础上,提出了基于两种不同原理的深度图粗配准方法。第一种方法中首先提出一种新的基于多分辨最小主曲率的网格角点检测算法,通过该算法直接提取网格曲面上的角点特征,并查找两幅深度图间相匹配的角点,利用最小二乘法实现深度图的粗配准。第二种方法根据深度图数据的特点,利用深度图中每个空间点的形状指数,将深度图映射为二维合成灰度图像,并利用二维图像领域的特征检测与匹配技术提取合成灰度图像上的特征点,再通过映射关系间接地在深度图网格上提取三维匹配特征点对,最后通过进一步去除误匹配的优化策略实现深度图粗配准。大量实验表明,两种粗配准算法对噪声和数据重叠度均具有较高的鲁棒性,特别地,当深度图上几何特征较少时,第二种方法表现出更高的稳定性。3、深入研究了多视角测量数据的精配准理论和方法,对基于虚拟弹簧力系统的全局精配准算法提出了改进。主要的改进之处有两点:一是克服了原算法中不存在外点的假设条件,在最近对应点查找算法中增加边界约束,并依据迭代过程中的配准误差自动设置相应的权重,提高了配准精度;二是针对算法计算效率较低的问题,采用基于GPU的高效并行加速技术,明显提高了对应点的搜索效率。4、提出了一种基于网格融合的多视点深度图同时融合算法。算法依次将每幅深度图定义为基准图,在基准图基础上同时对其它多幅深度图重叠区域进行检测和优化调整,有效降低了融合累积误差。为使去除冗余后的基准图之间的缝隙能够光滑连接,融合算法中提出了边界重叠对应点约束方法以及设置层次边界加权方法,在缝隙缝合过程中综合考虑了数据的重叠与删除信息,简化了缝合难度,使得无需重新三角化或增加新点即可缝合多视图之间的缝隙,形成单一拓扑流形的融合网格曲面。在实现几何数据融合的同时,提出了一种纹理融合方法,能够在几何数据融合的统一框架下得到平滑过渡的纹理融合结果。多个实际测量模型的融合实验结果验证了本文融合算法有很好的效果。5、提出了一种具有各向异性的混合滤波算法。该算法定义了一种基于二阶邻域面的双边滤波算子,并通过分析比较该算子与基于一阶邻域点的双边滤波算子对光顺噪声和保持特征的不同原理,将这两种算子通过自适应设置的权重进行加权合并得到新的混合滤波算子,新算子可以在对噪声进行有效光顺的同时对光滑区域起到很好的保持作用,并有效抑制了多次光顺迭代所产生的网格模型收缩或扩张。6、提出了一种二次误差测度(QEM)网格简化法的改进算法。算法提出了网格简化的支撑域概念,在简化过程中通过查找折叠边在初始网格上的支撑域,建立简化边与初始网格之间的联系。在折叠过程中增加计算折叠边的新顶点到初始网格上相应支撑域的全局误差,并将此误差引入到QEM的误差测度中。实验表明,在相同简化率下,改进的简化算法比原始QEM算法较为明显地降低了简化误差,同时更好地保持了初始网格的细节特征。更多还原

【Abstract】 Digital model construction by measuring the shape of real-world objects has been widely used in an extensive range of fields, such as aerospace, automotive, shipbuilding, machinery manufacturing, biomedical, gaming and entertainment, etc. As a bridge built between the data acquisition and the data application, data processing is an important part of the 3D shape measurement. With the rapid development of digital imaging and image processing technology, 3D measurement systems with area-array cameras as the main sensors have made great progress in recent years. Several key techniques of data processing in this type of systems are studied in this paper, including data preprocessing, crude registration and fine registration of multi-view range images, range images integration, mesh smoothing and simplification. The main contents and innovations are summarized as follows.By utilizing both the depth information and the pixel structure information, a dense point clouds triangulation method and an adaptive sampling triangulation method are presented respectively. With the dense point clouds triangulation method, a full resolution mesh model of the range image can be quickly constructed. Whereas the adaptive sampling triangulation method, which is designed for real-time display of the massive measurement data, generates a mesh model with high appearance fidelity from significantly reduced points.Two crude registration methods are put forward based on two different principles. In the first one, the feature points are extracted directly from the 3D meshes by a new corner detection algorithm proposed. Having established the correspondences between two range images, the crude registration is achieved by the least square method. In the second one, the range image is mapped into a 2D artificial intensity image according to the shape index of every point in the range image. By using feature detection and matching methods for intensity image, feature points in the artificial image are extracted and matched, and then the 3D point matches are obtained indirectly by the mapping. After further steps for removing mismatches, the crude registration is carried out finally. Experiments show that both proposed methods are robust against the overlapping extent and the noise. The second method shows higher stability to register the range images with less geometric features.The theories and methods of fine registration of multi-view measurement data are deeply studied. In particular, an algorithm is proposed to improve the global fine registration which is based on the virtual spring force. There are two major improvements. First, to overcome the assumption in the original algorithm that the external points do not exist, the boundary constraint is imposed on the process searching for the nearest corresponding points. The corresponding weights are adaptively set according to the registration error in each iteration to improve the registration accuracy. Besides, the parallel acceleration technology based on GPU is adopted to improve the efficiency of the correspondance searching process.An algorithm for simultaneously merging the registered range images is proposed. Each range image is in turn defined as the benchmark image, and on this basis, the overlapping areas between the benchmark with the other range images are detected simultaneously and adjusted optimally to reduce the accumulated error. To connect the gaps smoothly between the images after redundancy removal, an overlapping constraint as well as a weighing scheme for the boundary points is presented. The overlapping and removal information are took into account in the stitching process, which simplifies the stitching algorithm, and a complete topological manifold mesh surface can be generated without re-triangulation or additional new points. Together with the geometric data integration, a texture blending method is presented to generate smooth texture fusion under the same framework. Experiments demonstrate the validity and efficiency of the proposed algorithm.A hybrid filtering algorithm with anisotropic is proposed. The bilateral filtering operator based on two-order neighborhood faces is defined and compared with another operator based on one-order neighborhood points in detail, The different performance of the two operators in smoothing noise and preserving features is considered, and a new hybrid filtering operator is proposed by combining the two operators with adaptive weights. The new operator is effective in filtering noise while keeping the smooth areas little changed. It is also superior in suppressing the mesh contraction or expansion that could be available after lots of iterations.An improvement to the quadric-error-metrics(QEM)-based mesh simplification algorithm is presented. The support region on the original mesh is defined and searched for the every collapse edge in the simplifying process. The connection between the collapse edge and the original mesh is accordingly established. The quadric error from the new vertex of the collapse edge to its support region is calculated as the global simplification error and is introduced into the total cost function of QEM. Experimental results demonstrate that the improved algorithm reduces the simplification error obviously and preserves the original mesh details better.更多还原