最新的和准确的工业场地三维模型需要不同的应用，如规划、文件和培训。传统的获取竣工信息的方法，如通过磁带和视距仪进行手动测量，不仅速度慢且繁琐，而且大多数情况下，它们也无法提供所需的详细信息。由于存在放射性、有毒或有害物质以及不安全的工作环境，许多工业设施为人员提供了有限的通道，因此必须使用非接触测量方法。传统的摄影测量依赖于点或线的测量，如果不进行大量的手工编辑和细化，很难得到完整的CAD模型。与摄影测量相比，激光扫描提供了清晰而密集的三维测量。在过去的十年里，激光扫描仪的速度和精度都有了快速的提高，同时其成本和尺寸也在不断的缩小。对于大多数建模任务，市场上所有可用的建模工具都依赖于大量的操作员干预。虽然有一些半自动工具，如平面或圆柱体生长，即使如此，操作员必须针对每个基本体的生长过程。此外，装配面必须由操作员手动编辑，以将其转换为CAD描述。本文提出了一种新的方法和技术，可用于从点云和图像中自动或高效地半自动地对现有工业装置进行三维建模。目标是利用点云的显式三维信息来自动检测场景中的物体和结构，然后将检测到的目标作为基于模型的配准目标，通过搜索目标的对应关系实现配准的自动化。为了避免手工编辑，所提出的技术使用从一个常见的CAD对象目录中的模型作为模型拟合的模板。在最后的拟合过程中还加入了相位图像，提高了参数估计的质量。

分割是目标识别和模型拟合中一个非常重要的步骤。提出了一种点云分割方法，在将输入数据分割成相互不相交、平滑连接的区域时，避免了过度分割。它采用了一种基于曲面法向相似性和空间连通性相结合的准则，我们称之为平滑度约束。由于不使用曲面曲率，我们的算法对噪声不太敏感。此外，只有少数参数可以调整，以在欠分割和过分割之间获得所需的折衷。

分割之后是基于Hough变换的目标识别阶段，用于在点云中自动检测平面和圆柱体。对于平面检测，Hough变换是三维的。对于圆柱的检测，直接应用Hough变换需要一个5D的Hough空间，由于其空间和计算复杂性，这是不现实的。为了解决这个问题，我们提出了一个两步的方法，需要一个二维和三维霍夫空间。在第一步中，我们检测到圆柱体方向的强假设。第二步估计圆柱体的其余三个参数，即半径和位置。

平面、圆柱体、球体、圆锥体、圆环体和CSG等模型与点云的拟合问题是数据约简的一个重要问题。对于CSG模型的拟合，本文提出了三种不同的正交距离逼近方法，并从速度和精度两个方面进行了比较。我们还提出了在单个扫描中使用建模对象作为目标进行配准的方法。由于使用了现有的几何结构，因此不需要放置人工目标。为此，我们提出了两种不同的方法，称为间接法和直接法。间接法是一种快速获得近似值的方法，而直接法则用于细化近似解。我们还提出了自动找到相应扫描对象配准的技术。所提出的技术是基于约束传播的，它利用先前的对应决策中的几何信息来过滤出未来对应的可能性。

尽管点云由于其明确的三维信息而对自动化非常重要，但是图像提供了一个互补的信息源，因为它们包含边界对象的明确定义的边缘。我们提出拟合CSG模型的方法结合了点云和图像。我们还提出了技术规范之间的几何约束子部分的CSG和它们的模型估计过程，给出了常见几何约束的分类及其数学表达式。我们希望本论文所提出的技术能够提高从点云和图像中获取工业装置模型的效率和质量。

Up to date and accurate 3D models of industrial sites are required for different applications like planning, documentation and training. Traditional methods for acquiring as-built information like manual measurements by tape and tacheometry are not only slow and cumbersome but most of the time they also fail to provide the amount of detail required. Many industrial facilities provide a limited personnel access because of the presence of radioactive, toxic or hazardous materials together with an unsafe working environment, which necessitates the use of non-contact measurement methods. Traditional photogrammetry depends on point or line measurements from which it is very hard to get complete CAD models without extensive manual editing and refinement. Compared to photogrammetry laser scanning provides explicit and dense 3D measurements. There has been a rapid increase in the speed and accuracy of the laser scanners in the last decade, while their costs and sizes have been continuously shrinking. All modeling tools available on the market depend on heavy operator intervention for most of the modeling tasks. Although there are some semi-automatic tools like plane or cylinder growing even there the operator has to start the growing process for each primitive. Furthermore, the fitted surfaces must be manually edited by the operator to convert them to a CAD description. This thesis presents new methods and techniques which can be used for automatic or efficient semiautomatic 3D modeling of existing industrial installations from point clouds and images. The goal is to use explicit 3D information from the point clouds to automatically detect the objects and structure present in the scene. The detected objects are then used as targets for model based registration, which can be automated by searching for object correspondences. To avoid manual editing the presented techniques use models from a catalogue of commonly found CAD objects as templates for model fitting. In the final fitting phase images are also included to improve the quality of parameter estimation.

Segmentation is a very important step that needs to be carried out as a pre cursor to object recognition and model fitting. We present a method for the segmentation of the point clouds, which avoids over-segmentation while partitioning the input data into mutually disjoint, smoothly connected regions. It uses a criterion based on a combination of surface normal similarity and spatial connectivity, which we call smoothness constraint. As we do not use surface curvature our algorithm is less sensitive to noise. Moreover, there are only a few parameters which can be adjusted to get a desired trade-off between under- and over-segmentation.

Segmentation is followed by a stage of object recognition based on a variation of the Hough transform for automatic plane and cylinder detection in the point clouds. For plane detection the Hough transform is three dimensional. For the cylinder detection the direct application of the Hough transform requires a 5D Hough space, which is quite impractical because of its space and computational complexity. To resolve this problem we present a two-step approach requiring a 2D and 3D Hough space. In the first step we detect strong hypotheses for the cylinder orientation. The second step estimates the remaining three parameters of the cylinder i.e. radius and position.

The problem of fitting models like planes, cylinders, spheres, cones, tori and CSG models to point clouds is very important for data reduction. For the fitting of CSG models this thesis presents three different methods for approximating the orthogonal distance, which are compared based on speed and accuracy. We also present methods for using modeled objects in individual scans as targets for registration. As the available geometric structure is used, there is no need to place artificial targets. We present two different methods for this purpose called Indirect and Direct method. The Indirect method is a quick way to get approximate values while the Direct method is then used to refine the approximate solution. We also present techniques for automatically finding the corresponding objects for registration of scans. The presented techniques are based on constraint propagation which use the geometric information available from the previously made correspondence decision to filter out the possibilities for future correspondences.

Although point clouds are very important for the automation because of their explicit 3D information, images provide a complementary source of information as they contain well-defined edges of the bounded objects. We present methods for the fitting of CSG models to a combination of point clouds and images. We also present techniques for the specification of geometric constraints between sub-parts of a CSG tree and their inclusion in the model estimation process. A taxonomy of commonly encountered geometric constraints and their mathematical formulation is also given.

We hope that the techniques presented in this thesis will lead to an improvement in efficiency and quality of the models obtained for industrial installations from point clouds and images.

1.       引言

2. 基于平滑约束的分割

3. 目标识别

4. 点云模型拟合

5. 基于模型的配准

6. 自动化搜索

7. 约束的CSG拟合

8. 结论

附录方向空间的均匀采样

更多精彩文章请关注公众号：