CASIA OpenIR  > 毕业生  > 博士学位论文
基于图像的城市场景三维模型重建
隋伟
学位类型工学博士
导师潘春洪
2016-05-30
学位授予单位中国科学院研究生院
学位授予地点北京
学位专业模式识别与智能系统
关键词城市数字化 基于图像的模型重建 结构保持的模型重建 运动恢复 结构 多视角立体技术 图拉普拉斯 平面分割 表面重建 纹理映射
其他摘要    随着硬件显示设备和虚拟现实技术的发展,具有真实感的数字化城市受到 人们越来越多的关注。数字化城市在城市规划、三维地图、古迹保护和虚拟仿 真等方面发挥着重要的作用,并有巨大的应用潜力和市场价值。建筑物是城市 场景中重要的组成部分,也是数字化城市的主要对象。如何有效地重建建筑物 的三维模型一直是数字化城市的研究热点。
 缺陷的点云中恢复完基于图像的模型重建方法由于其数据获取的便捷性、经济性和高效性,为 数字化城市提供了一条有效的途径。基于图像的三维模型重建技术是一项系统 工程,它包括基于图像的三维点云获取,几何结构的恢复和场景的绘制等过 程。随着计算机视觉技术和图形学技术的发展,人们已经可以从图像中获取稠 密的点云数据。具有高度真实感的场景绘制技术也已经较为成熟,但是如何从 点云中获取建筑物的三维模型这一问题仍然没有得到解决。城市建筑模型重建 的难度主要存在于三个方面:(1) 建筑中通常含有很多的尖锐结构(如线段、平 面等),如何在模型重建过程中保持这些尖锐结构是建筑模型重建的难点之一;
(2) 城市建筑的结构通常具有高度复杂性和多样性,如何通过数学建模的方法 对建筑模型进行表述是建筑模型重建的难点之二;(3) 基于图像的方法获取的 点云通常含有严重的缺陷,例如噪声和缺失等。 如何从有缺陷的点云中恢复完 整、准确和简洁的建筑模型是建筑模型重建的难点之三,同时这也一直是计算 机视觉和计算机图形学领域一个重要的研究课题。本文围绕上述问题开展相关 研究,主要贡献如下:

  1 提出了一种基于图拉普拉斯正则的平面分割模型,并将其用于城市建筑 点云的平面元素提取。该方法在全局上采用一种线性投影模型进行分 片平面拟合,在局部上采用图拉普拉斯模型保持局部的一致性。图拉普 拉斯有效地整合了建筑的多种先验信息,从而显著地提高了分割的准 确率。该模型可以通过乘性更新的方法快速进行优化。与已有的方法相 比,本文的方法能够同时处理多个平面结构且对参数不敏感。真实数据 和仿真数据上的实验结果表明,该模型对噪声鲁棒,具有较强的平面分割能力。
  2 提出了一种基于断层的全自动城市建筑模型重建算法,用于从点云中恢 复建筑的三维模型。该方法的核心思想是首先将建筑沿垂直于地面的方 向切割成一系列断层,然后在每一片断层上自动重建分段线段的平面轮 廓并从中提取主导平面轮廓。之后将主导平面轮廓作为标签通过马尔科 夫随机场传播到其它断层中。这样建筑的三维模型重建问题便转化成了 基于马尔科夫的多标签分配问题。为了准确地提取和传播主导平面轮 廓,本文提出了一种基于语法编码的方法来准确地度量平面轮廓之间的 相似性。此外,为了恢复建筑模型表面的窗户结构,本文提出了一种基 于图像的模型编辑方法,该方法对于场景中重复出现的窗户结构十分有 效。本章的重建算法适用于分段平面且与地面垂直的建筑,其主要优点 是能够保持建筑的尖锐结构,同时生成简洁、水密的模模型。大量真实数 据上的实验结果验证了本文算法的有效性。
  3 构建了一套实用的基于图像的大规模城市场景模型重建系统。该系统在 集成了基于图像的点云生成技术和由点云到模型的重建技术的基础之 上,实现了全自动的纹理图像拼接与编辑功能。该系统输入多视角图像 序列,自动输出真实感较强的纹理模型。由点云到模型的重建直接影响 到纹理模型的真实感,针对城市场景航拍图像得到的点云数据,本文对 几种常用的表面重建算法的性能进行了测试和评估,从而为城市场景的 模型重建提供参考。该系统能够对较大规模的城市场景进行自动重建, 并支持测绘学中矢量数据的交互式采集和输出。

;     With the development of display technology and virtual reality technology, realistic digital city has been attracting more and more attention in the world. Digital city plays an important role in applications such as urban planning, 3D maps, heritage protection and virtual simulation, etc., and are also of high economical value and market potentiality. As the major component of urban scene, buildings are the main object of city digitalization. Hence, how to effectively recover 3D building models has being a hot topic for city digitalization.
     Image based modelling has provided an effective way for city digitalization, for the high convenience, economy and  efficiency of its data acquisition. Image based modelling is indeed a systematic project containing image based data acquisition, geometry recovery and model rendering. Due to the development of computer vision and computer graphics, people have been able to obtain dense point clouds from images. There also have been mature technologies for realistic model rendering. However, the problem of recovering architecture geometry from point clouds still remains unresolved. The main difficulties of urban building reconstruction lie in three aspects. First, urban buildings are commonly composed of sharp features, such as polygons and edges. How to preserve these sharp features during the reconstruction is a challengingproblem. Second, urban building styles are always of high variety and complexity,which means it is hard to describe building models via uniform mathematical models. Third, the data acquired from image based methods are unstructured and usually have severe defects, such as noisy and large area missing. How to recover complete, compact and accurate building models from imperfect point clouds is another challenging problem and also has been an active research topic in computer vision and computer graphics community. In this dissertation, we have done solid research on the urban building reconstruction problem, and our innovative work are including:
     We proposed a novel graph Laplacian regularized K-planes model for extracting planes from urban building point clouds. In this method, a linear projection model is utilized to fit planar surfaces globally while a graph Laplacian regularization is applied to preserve smoothness of each plane locally. The graph Laplacian regularization has effectively incorporated multiple sources of prior information which improves the segmentation accuracy apparently. The objective function can be fast minimized via an iterative updating algorithm. Compared with the state of arts, our method is able to extracting multiple planes simultaneously while keeps free of parameter tuning. Comparative experimental results on both synthetic and real data sets demonstrate that the proposed method has good robustness to noise and strong ability of planar segmentation.
     A new layer-wise based automatic urban building modelling method is proposed to recover building geometry from 3D point clouds. In this method, point clouds are first cut into a sequence of slices along the gravity direction. After that piecewise linear floorplans are automatically reconstructed from all slices among which dominant floorplans are extracted. In the following dominant floorplans are propagated to other floors as labels via Markov Random Field (MRF) framework. Thus, the modelling problem is converted into an MRF based multi-label assignment problem. To guarantee the accuracy of dominant floorplan extraction and propagation, we have proposed a grammar based method to measure similarity between floorplan shapes.  In addition, an image based model editing method is designed to recover detailed window structures which is quite effective when window structures appear repetitively. The proposed method is suitable for buildings composed of planar polygons and aligned with the gravity direction, which are common in urban area. The main advantages of the proposed method are that it achieves feature preserving while generating compact and watertight models. Experimental results on both synthetic and real data sets have validated the effectiveness of our method.
   A practical image based modelling system is developed for large urban scene reconstruction. Combined with the image based data acquisition and surface reconstruction techniques, the system has implemented multi-view based texture image stitching and editing algorithm. Realistic texture models can be generated from this system with full automation. In addition, to provide reliable reference for surface reconstruction, we have made evaluations of several methods' performances on urban point clouds. The system is able to reconstruct a large scale scene with thousands of images. Multiple reconstruction examples have demonstrated the effectiveness of our system.

学科领域模式识别与智能系统
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/11701
专题毕业生_博士学位论文
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
隋伟. 基于图像的城市场景三维模型重建[D]. 北京. 中国科学院研究生院,2016.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
Thesis_new_version.p(54849KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[隋伟]的文章
百度学术
百度学术中相似的文章
[隋伟]的文章
必应学术
必应学术中相似的文章
[隋伟]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。