Rekonstruktion von Fassadenstrukturen mittels formaler Grammatiken und Reversible Jump Markov Chain Monte Carlo Sampling
Rekonstruktion von Fassadenstrukturen mittels formaler Grammatiken und Reversible Jump Markov Chain Monte Carlo Sampling
复制标题
形式化语法和可逆跳转马尔可夫链蒙特卡罗采样的法萨登结构重构
DOI：
--
复制
发表时间：
2010
期刊：
影响因子：
--
通讯作者：
Nora Ripperda
中科院分区：
文献类型：
--
作者：
Nora Ripperda
三维建筑物模型用于各种应用中。这些可以在旅游，城市规划或3D导航等领域找到。随着应用数量的不断增加，对3D模型的需求也在不断增长。为了满足要求并保持数据最新，需要自动重建方法。由于对细节层次的要求越来越高，本文提出了一种自动立面重建的方法。对于重建图像和深度数据，使用地面激光扫描仪采集。本文提出了一种新的方法，将可逆跳马尔可夫链蒙特卡罗方法与形式文法相结合。使用的语法规则和先验，这是来自相对频率，确保结构和随机属性的外墙被认为是同样的。重建方法特别强调对立面结构的分析。分析立面图像以获得关于立面结构的信息。接收到的信息被用作重建过程中的先验知识。关于立面的知识用形式语法来表述。它包含经常出现在立面上的图案，如网格结构中的窗户，对称或重复。如果在数据中检测到这些结构，则它们可以用于数据的紧凑存储或用于概括。本论文的第二个重要议题是自动重建程序的发展。该方法使用facade语法规则自动生成最适合数据的派生树。本文采用了可逆跳马尔可夫链蒙特卡罗方法。这个随机过程提出了一个变化的马尔可夫链根据规则的门面语法。根据接受概率，变更将被接受或拒绝。此外，本文还讨论了接收概率的确定问题。重要的是要确保模型和数据的拟合质量与模型复杂性之间的平衡。为了考虑到这一点的评分功能的基础上最小的描述长度的开发。最后对重建结果进行了分析。使用由深度和图像数据组成的六个数据集和仅由深度数据组成的一个数据集对重建方法进行了测试。七个立面中有六个在结构和立面元素的位置上得到了正确的重建。所有重建的正确性和完整性在95.3%和88.3%之间。此外，外观语法的表现力进行了测试的外观图像数据库，其中包含56个图像的援助。立面重建正确率为33.9%，错误率为21.4%。剩下的45%与立面相比有很小的变化。第一组的正确率在94.1%~ 84.6%之间，完整率在93.9%~ 81.8%之间。
Three-dimensional building models are used in a variety of applications. These can be found in fields such as tourism, city planning, or 3D navigation. Due to the increasing number of applications, the demand on 3D models is growing. In order to cover the requirements and to keep the data up to date, automatic reconstruction methods are needed. Because the requirements in the level of detail are increasing simultaneously, in this thesis a method for automatic facade reconstruction is developed. For the reconstruction image and depth data are used, which were acquired using a terrestrial laser scanner. In this thesis a new method is developed, which combines a Reversible jump Markov Chain Monte Carlo method with formal grammars. The use of grammar rules and priors, which are derived from relative frequencies, ensures that structural and stochastic attributes of facades are considered likewise. The reconstruction method places special emphasis on the analysis of facade structures. Facade images are analysed to obtain information about the structure of facades. The received information is employed as prior knowledge in the reconstruction process. The knowledge about facades is formulated in a formal grammar. It contains patterns which occur frequently on facades like windows in a grid structure, symmetries or repetitions. If these structures are detected in the data, they can be used for compact storage of the data or for generalisation. The second important issue of this thesis is the development of an automatic reconstruction procedure. This method generates the derivation tree automatically, which fits the data best, using the rules of the facade grammar. In this thesis a Reversible jump Markov Chain Monte Carlo approach is used. This stochastic process proposes a change of the Markov Chain according to the rules of the facade grammar. Depending on an acceptance probability the change will be accepted or rejected. Furthermore, this thesis addresses the determination of the acceptance probability. It is important to ensure a balance between the quality of the fit of model and data and the model complexity. To take this into account a scoring function based on minimum description length is developed. Finally the results of the reconstruction are analysed. The reconstruction method was tested with six data sets consisting of depth and image data and one data set consisting of depth data only. Six of seven facades were reconstructed correctly in structure and position of the facade elements. The correctness and completeness of all reconstructions lies between 95.3% and 88.3%. Additionally the expressiveness of the facade grammar was tested with the aid of a facade image database which contains 56 images. 33.9% of the facades were reconstructed correctly, 21.4% were reconstructed wrong. The remaining 45% show small variations compared to the facade. The correctness of the first group lies between 94.1% and 84.6% and the completeness between 93.9% and 81.8%.