基于影像的近景目标三维重建若干关键技术研究

发布时间：2018-03-29 01:36

本文选题：近景目标　切入点：三维重建　出处：《武汉大学》2014年博士论文

【摘要】：近景目标的三维重建是数字摄影测量与计算机视觉领域中一项复杂的综合性技术,可用于国民生产各行各业。本文选取了基于结构光的三维重建技术、基于轮廓线的三维重建技术、几何模型与多视影像配准三个典型问题中的若干关键技术进行了深入研究,具体研究内容和研究结果如下： (1)单帧投影下的结构光扫描关键技术。本文完成了伪随机投影图案的设计和特性分析,邻域窗口沿核线方向的全局唯一性分析,提出了一种SEEM快速影像匹配方法,包括核线影像种子点匹配和基于区域增长的密集匹配,进行了错点剔除和匹配算法并行优化等。通过实验数据提出了两个长度为n的随机序列的相关系数服从均值为0、标准差为1/(?)n的正态分布的假设,在此假设下,使用11×11的窗口进行种子点匹配可以在整条核线上保证匹配点的唯一性,使用5×5的窗口进行区域增长匹配可以在小范围内保证匹配点的唯一性。一旦出现了错误的匹配,根据伪随机投影图案的性质,区域增长算法将迅速达到边界,这是从匹配点中剔除错点的关键。在扫描距离约为600mm,基线长度约为276mm时,点云的绝对精度为0.185mm,相对精度为3200分之一,像方精度约为0.3像素。在CPU Core2T5850、内存2GB的配置下,算法达到了每秒约40万点的匹配速度。 (2)轮廓线约束下的多视影像三维重建。本文完成了基于图割的物体轮廓线半自动提取,轮廓线的平滑,轮廓线图像的四叉树森林压缩存储,体素模型和表面网格模型的生成算法,表面网格模型的优化。不同的赋权方式会得到不同的图割结果,但没有哪一种赋权方式具有明显的优势。为了尽可能地减少人工交互的工作量,在拍摄影像时应尽量选择与物体的亮度、颜色相差较大的背景。由于在生成体素模型和表面网格模型时,需要将所有的轮廓线图像同时读入内存中(每张影像约10MB),采用四叉树森林存储轮廓线图像这种近似二值图像,可以达到100到300倍的压缩率,而访问四叉树的叶子节点需要从根节点开始逐层向下搜索,最好情况是1层,最坏的情况是9层。对于圆形过渡的曲面,拍摄影像的角度的任意性对其重建的主观效果影响不大,适当增加影像的数量就可以得到更精细的结果；对于较大的平面,很容易形成三角形的凸起,只有当某一张影像的摄影中心刚好位于或者接近该平面时,才可以消除这种现象,拍摄时需针对性地选择视角；对于存在凹陷的部分,本章节的方法从理论和实践上都无法得到好的重建结果；本章节中用于实验的物体表面都缺乏(或者局部缺乏)纹理,无法使用影像匹配的方法得到更好的重建结果,是在不使用结构光扫描(或其他相似的技术手段)前提下,本章节的方法是最合适的。 (3)几何模型与多视影像配准。本文完成了几何模型与多视影像的粗配准(空间相似变换),基于互信息的精配准,包括OpenGL渲染图的生成方法,渲染图上几何特征显著区域的选择,灰度联合直方图的统计,基于Powell方法的配准参数优化等。在粗配准中,需要首先根据物体和背景的相对关系将影像分成不同的组,每一个影像组位于同一个坐标系中。影像组坐标系与几何模型坐标系之间、不同的影像组坐标系系之间存在不同的配准传递策略,在人工选取同名点时,可以根据实际情况选择不同的策略,最终将所有的影像组坐标系都配准到几何模型坐标系中。在精配准中,OpenGL光照方向的不同、直方图统计区域的选择都会影响最终的配准结果,光照方向与真实影像的光照条件越接近,配准精度越高,只统计几何特征的影响大于纹理特征影响的区域,将得到更好的配准结果。对与具有镜面反射特性的物体,由于影像的灰度依赖于物体的几何曲率变化,故经过基于互信息的优化后,配准参数的精度得到提高。在一个影像组仅有1张影像的特殊情况下,舍去缩放系数,仅对6个参数进行优化,仍可以得到较好的配准结果。基于空间相似变换的粗配准需要少量的人工交互,每一个影像组需要至少3对同名点,每组耗时在5分钟之内；而基于互信息的精配准的耗时也在可接受的范围之内。
[Abstract]:The 3D reconstruction is close range digital photogrammetry and computer vision in the field of a complex technology, can be used in all walks of life of national production. This paper selects the technology of 3D reconstruction based on structured light, 3D reconstruction based on contour lines, some key technologies of the three typical problems of geometric model and multi view image registration in the the in-depth research, specific research contents and results are as follows:
(1) the key technologies of optical scanning structure of single frame projection. This paper has completed the design and analysis of characteristics of pseudo random projection patterns, analysis of the uniqueness of the global neighborhood window along the epipolar direction, SEEM presents a fast image matching method, including dense matching epipolar images based on point matching and seed region growing that was the wrong point detection and matching algorithm and parallel optimization. Through the experimental data presented a correlation coefficient of two with a length of N random sequence with mean 0 and standard deviation of 1/ (?) n normality assumption. Under this assumption, the seed point matching can ensure the uniqueness of the matching points in the entire nuclear line using 11 x 11 window for regional growth, matching can ensure the uniqueness of the matching points in a small range using 5 x 5 window. Once the wrong matching, according to the properties of pseudo random projection pattern, region growth algorithm Will quickly reach the boundary, which is the key to eliminate the wrong matching points. From the point of the scanning distance is about 600mm, the baseline length is about 276mm, the absolute accuracy of point cloud is 0.185mm, the relative accuracy of 3200 quarter, as the accuracy is about 0.3 pixels. In CPU Core2T5850, 2GB memory configuration and the algorithm to achieve the matching speed of about 400 thousand points per second.
(2) contour under the constraints of multi view images of three-dimensional reconstruction. The graph cut object contour extraction based on semi automatic, smooth contour, contour image compression and storage of four forest tree generation algorithm of voxel model and surface mesh model, optimization of the surface mesh model. Different weighting the way to get a different graph cut results, but there is no a weighting approach has obvious advantages. In order to reduce the workload of manual interaction, the image should be chosen as the brightness of the object, the color difference between the larger background. Due to the voxel model and surface mesh model, need to profile line all the image read into memory at the same time (in each image, the image is about 10MB) four tree forest storage contour approximate two value image, can achieve a compression ratio of 100 to 300 times, and access to the four fork tree leaf Festival To start from the root node layer down search, the best is 1, the worst case is the 9 layer. Surface for the circular transition, has little effect on arbitrary image angle on the subjective effect of the reconstruction, increasing the number of images can get more precise results for larger; the plane, it is easy to form the triangle convex, only when a picture of the photography center is located in or near the plane, can eliminate this phenomenon, need to choose from when shooting; for the existence of concave part, method of this chapter from both theory and practice to get good reconstruction results; for the surface experiment are lacking in this chapter (or the lack of local texture), image matching methods cannot be used to get better reconstruction results, is not in use structured light scanning (or other similar technology On the premise of this section, the method in this section is the most appropriate.
(3) geometric model and multi view image registration. The geometric model and multi view image registration (space similarity transformation), the fine registration based on mutual information, including the method of generating OpenGL rendering, geometry rendering salient region selection, combined with gray histogram statistics, Powell method the registration parameters optimization. Based on Rough Registration, need first to divide the image into different groups according to the relation between the object and background, each image group is located in the same coordinate system. Between the image coordinates and the geometric model of group coordinate system, there are different registration transfer strategy between image coordinates the same point in the artificial selection, you can choose a different strategy according to the actual situation, will all the image coordinate system are registered into the geometric model of coordinate system. In the fine registration, OpenGL light with different illumination direction, straight The square of the regional statistical selection will affect the final registration result, the direction of illumination and real image illumination conditions closer, registration accuracy is high, influence only statistical geometrical features than the texture features of the affected area, will get better registration results. And with specular reflection objects due to geometric curvature change of image gray dependent objects, so the optimized based on mutual information, the registration parameters precision has been improved. In a special case of image group only 1 images, to optimize the zoom coefficient, only 6 parameters, you can still get a better registration results. The coarse registration space similarity transformation the need for manual interaction based on a small number of images, each group of at least 3 points, each time in 5 minutes; and based on the time-consuming fine registration of mutual information can be accepted within.

【学位授予单位】：武汉大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：P234.1

【参考文献】