利用倾斜影像进行三维数字城市重建的关键技术研究

发布时间：2018-06-02 21:06

本文选题：倾斜影像 + 倾斜影像自动空中三角测量　；参考：《武汉大学》2014年博士论文

【摘要】：倾斜摄影测量技术是当前的研究热点之一。相对于传统的垂直航空摄影,倾斜航空摄影具备有效地获取地物侧面纹理、可对同一地区进行多视角重复观测,能获得更符合人眼视觉的直观世界的影像等优势,因而被广泛应用于数字三维城市建模,城市管理与规划,地图服务等领域。如今机载倾斜多相机系统已经成为快速发展的空间信息市场的标准技术,世界上大量的科技公司(Blom Oblique, IGI, Leica, Midas, Pictometry, Vexcel/Microsoft, VisionMap)都生产了倾斜多相机系统。另一方面,近年来市场对三维城市模型的需求大量增加,世界各地的生产商致力于提供数字城市三维模型。微软,Google等大型科技公司更是着手建立世界范围内数以千万计的城市三维模型。工程浩大,若使用传统的人工方式建立城市三维模型将带来高昂的投入。因而需要一种高度自动化的数字三维城市重建技术。由于倾斜摄影技术在获取顶面纹理的同时,其搭载的倾斜相机能够同时获得地物的侧面纹理,具备了传统的航摄相片和地面影像的双重优势,因而是数字城市三维重建的首选技术手段。但是相对于硬件的蓬勃发展,倾斜影像的处理软件的发展相对滞后。主要原因在于倾斜视角的存在,导致获得的相片遮挡严重,且相片内部的尺度不一,造成了倾斜影像数据处理上存在一定的困难。传统的摄影测量软件由于设计之初只考虑了垂直摄影的情况,航片的倾角都不大,因而处理工艺上无法处理内容差异巨大的倾斜影像。可以看到虽然市场上对三维数字城市的需求量巨大,且倾斜影像是数字城市重建重要且是首选的数据源,但是利用倾斜影像进行自动化处理获得数字城市模型的技术相对滞后。针对此问题,本文提出一套利用倾斜影像进行自动化三维重建的解决方案,并同时提出了一种在正射影像上进行立体测图的方法,对上一步获得的三维表面模型提供矢量数据和独立模型数据,以满足GIS系统和数字城市的需求。本文的自动化重建主要分为两大步。第一步,本文提出了一套全自动化的倾斜影像空中三角测量方法以解决倾斜影像的空中三角测量问题,获得倾斜影像的精确外方位元素。首先本文利用POS数据自动化寻找具备重叠的影像的方法,使得整个流程不需要人工进行航带排列,即可获得影像的连接关系；然后本文通过POS数据制作一种“虚拟水平相片”的方法,解决相片中尺度不一带来的特征点描述具备偏差的问题,成功解决了倾斜影像的连接点匹配问题,并同时采用Ransac的5点法相对定向和反向匹配策略剔除匹配粗差；为进一步提高平差结果的精度,并摆脱人工判断剔除空三连接点粗差的过程,本文提出一种多尺度的粗差探测技术,通过分析点云空间分布和密度,自动剔除连接点中残留的粗差。不同于经典的利用可靠性理论的粗差剔除策略。本文的方法能够发现位于核线附近的两度重叠的粗差,而这种粗差在可靠性理论里是不可定位不可发现的。经典的基于统计学的粗差探测策略能够处理中等和小粗差,然而对大粗差的处理效果不佳。本文方法可以综合处理小粗差和大粗差,且可以处理少量的簇状粗差。实验表明,利用本文方法能够有效找出空三连接点的粗差点,有效提高平差精度,使得像点残差中误差,在0.5像素之内,降低为剔除粗差前的1/3。本文使用两个对比实验,分别将ASIFT法与著名的VisualSFM中采用的方法对比验证了本文的方法。实验表明,本文的方法优于使用ASIFT做为特征点提取与匹配的结果,空三结果中像点残差中误差仅为ASIFT方法的1/3,且核线残差中误差仅为ASIFT方法的2/5,运行时间是ASIFT方法的1/12000。而与VisualSFM的方法对比结果表明,本文方法获得的同名点数为该方法的3.7倍,虽然像点残差中误差与其相当,都低于0.5个像素,但是同名点的核线残差中误差仅为VisualSFM的1/3。第二步,为获得高质量的地表模型,本文提出了一种自动化的从倾斜影像中生成表面模型的方法。为使得整个处理流程能够自动化,本文首先提出了一个自动选择最优匹配像对的方法；并根据空三的连接点信息自动获取待匹配影像自动获取重叠区；为适应倾斜影像的特点,本文对经典的Semi-Global Matching方法进行改进,主要在几个方面改进,首先利用空三的结果自动获取了影像的重叠区避免了人工给定重叠范围,然后利用Census匹配算法获得稳健的初始视差图,接着将经典的SGM方法的8个方向独立优化的方式集成在两次循环中进行,最后为了解决SGM方法消耗内存过大,以及经典的SGM方法无法分块处理大影像的问题,本文使用动态获取视差搜索范围的方式降低内存的使用效率,进而提高算法的适用性。通过与著名的PMVS算法进行对比实验表明本文方法在完整性和匹配效果上都优于PMVS算法。最后,为了能够高效的获取地物的矢量信息和建筑物模型,本文提出了一种新的基于正射影像的立体测图方法。不同于以往的立体测图的方法,本文的测图方法与平台无关,且只需要在二维模式下测图,适应于现存的大量2D-GIS系统。另外本文的方法在获得矢量的同时,能够同时获得地物的三维简单模型,使得能够快速构建简单城市三维环境。相比于倾斜影像自动重建而得的模型而言,测图得到的模型是单个地物模型,而前者则是整个地表模型。本文提出的倾斜影像处理流程完全自动化,并获得了较好的效果。且基于正射影像测图的方法与影像来源和平台无关极具潜力。
[Abstract]:Inclined photogrammetry is one of the hotspots of current research. Compared with the traditional vertical aerial photography, the oblique aerial photography can effectively obtain the texture of the surface of the ground objects. It can carry on the repeated observation of the same area and can obtain the image of the visual world which is more conforming to the human vision. Therefore, it is widely used in the digital three-dimensional city. City modeling, city management and planning, map service and other fields. Now the airborne tilt multi camera system has become the fast developing space information market standard technology. A large number of technology companies in the world (Blom Oblique, IGI, Leica, Midas, Pictometry, Vexcel/ Microsoft, VisionMap) all produce the inclined multi camera system.
On the other hand, in recent years, the demand for three dimensional urban models has increased greatly, and manufacturers around the world are committed to providing digital urban 3D models. Microsoft, Google and other large technology companies are starting to set up a world - wide three-dimensional model of cities. The dimensional model will bring high input. Therefore, a highly automated digital 3D urban reconstruction technology is needed.
Because the tilted camera can obtain the top surface texture, the tilt camera can obtain the side texture of the ground objects simultaneously. It has the dual advantages of the traditional aerial photograph and the ground image. Therefore, it is the first choice for the 3D reconstruction of the digital city. But relative to the vigorous development of the hardware, the processing software of the inclined image is used. The main reason lies in the existence of the tilt angle, which leads to the serious occlusion of the obtained photos and the different scales inside the photo, which causes some difficulties in the processing of the inclined image data. We cannot handle sloping images with huge difference in content.
It can be seen that although the demand for 3D digital cities is huge in the market, and the tilt image is an important and first choice data source for digital city reconstruction, the technology of obtaining digital city model by automatic processing with inclined images is relatively lagging. In addition, a method of stereo mapping on the Orthophoto Image is proposed, which provides vector data and independent model data for the 3D surface model obtained in the last step in order to meet the needs of the GIS system and the digital city.
The automatic reconstruction of this paper is divided into two major steps. In the first step, this paper proposes a fully automated set of oblique image aerial triangulation to solve the problem of aerial triangulation in the inclined image and obtain the exact external elements of the inclined image. Firstly, this paper uses the POS data automatization to find the overlapped images. The whole process does not need to be arranged manually, and the connection of images can be obtained. Then, a "virtual horizontal photo" method is made through POS data to solve the problem that the feature points with different scales in the photo are described with deviation, and the connection point matching problem of the inclined image is successfully solved, and Rans is adopted at the same time. In order to further improve the accuracy of the adjustment result and get rid of the rough difference between the three connection points of the space, the 5 point method of AC's relative orientation and the reverse matching strategy is eliminated. In this paper, a multi scale detection technique of coarse difference is proposed. By analyzing the spatial distribution and density of the point cloud, the residual gross error in the connection point is eliminated automatically. In the classical method of using reliability theory, this method can find the overlap of two degrees near the core line, and this coarse difference can not be found in the reliability theory. This method can deal with small gross errors and large gross errors in a comprehensive way, and can handle a small number of cluster rough differences. The experiment shows that this method can effectively find the rough difference of the three connection points of the empty three, effectively improve the accuracy of the adjustment, make the error in the image point residual, and reduce the 1/3. in the 0.5 pixels to eliminate the gross error before using two Contrast experiments show that the ASIFT method is compared with the method used in the famous VisualSFM respectively. The experiment shows that the method of this paper is superior to the result of using ASIFT as the feature point extraction and matching. The error of the image point residuals in the empty three result is only 1/3 of the ASIFT method, and the error of the kernel line residuals is only 2/5 of the ASIFT method. Time is the 1/12000. of the ASIFT method and the comparison with the VisualSFM method shows that the number of points of the same name obtained by this method is 3.7 times that of the method, although the error in the image point residuals is equal to 0.5 pixels, but the error of the same name point is only 1/3. of VisualSFM.
The second step, in order to obtain the high quality surface model, this paper proposes an automated method to generate the surface model from the tilted image. In order to automate the whole processing process, this paper first proposes a method to automatically select the optimal matching image pair, and automatically obtains the matched image according to the connection point information of the empty three. In order to adapt to the characteristics of the skew image, this paper improves the classical Semi-Global Matching method, mainly in several aspects. Firstly, using the result of empty three to automatically obtain the overlap area of the image to avoid the artificial given overlapping range, and then use the Census matching algorithm to obtain the robust initial parallax graph, and then, then the robust initial parallax graph is obtained. The 8 independent optimization methods of the classical SGM method are integrated in the two cycle. Finally, in order to solve the problem that the SGM method consumes too much memory and the classic SGM method can not block the large image, this paper uses the dynamic acquisition of parallax search scope to reduce the use efficiency of the memory, and then improves the application of the algorithm. Compared with the well-known PMVS algorithm, this method is better than the PMVS algorithm in terms of completeness and matching effect.
Finally, in order to obtain the vector information and the building model of the ground objects efficiently, a new stereo mapping method based on the Orthophoto Image is proposed in this paper. Different from the previous stereographic mapping method, the mapping method is independent of the platform, and it only needs to be measured in the two-dimensional pattern, and is adapted to a large number of existing 2D-GIS systems. At the same time, this method can obtain a three-dimensional simple model of the ground objects at the same time, so that the simple urban three-dimensional environment can be constructed quickly. Compared with the model obtained by the automatic reconstruction of the tilted image, the model obtained by the mapping is a single ground model, and the former is the whole surface model.
The slant image processing process proposed in this paper is fully automated and has achieved good results. And the method based on the orthophoto mapping has great potential to have nothing to do with the image source and platform.
【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】：P23

【相似文献】