Multiple Aerial Image Matching And Automatic Building Detection

As the application of GIS expands in land management traffic transportation electricity power telecommunication urban engineering and city planning, the 2D GIS is not enough for the needs any more. 3DGIS require more and more 3D urban geographic data, especially the 3D building data. How to acquire vast DTM and 3D building data quickly has Become a hot research area. Despite the EIDAR gains more and more attention, the aerial photograph are still the major data sources of urban mapping. Aerial photograph has been being portable of high effect of rich information, Photogrammetry is still now the main techniques for fast urban mapping. But it is a hard work to compile building date from stereo pair.Automatic extraction and reconstruction of urban scene from Anal has been a challenge problem for many years. It can be dated back to 19 century. And from then on, it had been the key problem. In the past 20 years, much research work has been done and many papers have been published. The development tendencies are:(1) From single image to multiple images. Vagueness can be removed with multiple images. The different side of an object can be seen from multiple images.(2) From gray information to color information, even the multiple spectrums of images. Color information is useful for image segmentation, making building detection easier.(3) Interaction between 2-D and 3-D processing.(4) From single image information grouping to grouping with multiple images(5) Imaging geometry, the knowledge of object and spatial reasoning are investigated more deeply. The spatial reasoning and semantic models make the building detection more reliable.(6) Building model develops from single rectangle to complex building of planes patches.(7) Multiple sources information integrating, such as with LID AR data or map data The aim of this thesis is to investigate the processing of multiple images, includingimage segmentation, grid matching and DSM generation, line matching and 3D line generation, plane surface detection and grouping, the components of building, least square adjustment of building model, etc.This research work is based on the idea of that most buildings can be modeled as plane surfaces, which connect each other. The processing is all based on Delaunay Triangulation, from image segmentation, region representation, line matching, plane surface construction, to entire building model.Delaunay Triangulation is a good data structure for divide the 2D planes. But most triangulation code does not consider the data quality. Usually, duplications of points can mess the construction. And on the other side, structure attributes are important to store varieties of objects. The solution is to expand the data structure to tolerate the duplication of points and attach a structure data pointer. The counter of duplication solves the problem and keeps transparent to user.The building model of plane surfaces can be represented naturally by TIN. If the database of building model can be built, it can be used to reconstruct the buildings in automatic extraction. A practical method based on TIN is discussed of digitizing and store the building models with multiple images.Traditionally, the split and merge algorithm of image segmentation is based on quad tree data structure. But it is not a convenient way to express the topography of regions, the line segments and other information. A new framework is put forward here. It is called "TIN based image segmentation and grouping". It tries to integrate edge information and area information, directed by perceptional grouping rule. The constrained triangle mesh is constructed with edge segments extracted by EDISON or other algorithm. So the initial segmentation is moderate. The results show this is a good idea.Elevation information (DSM) is important for building detection, for elevation is an invariant, which does not vary with the weather, the light and other things, expect the it is ruined by earthquake or other power.Most research supposes that there exists DSM, which can be used for extraction. In fact, multiple images provide a good data for DSM generation. DSM can be generated by image matching based on reliable correlation. The relaxation method is adopted here to optimize the matching with the elevation difference.For multiple images matching, we usually have MPGC by Grun and Baltsavias. It is extended from least square matching by Ackerman, which is for a pair of images. It supposes there are a group of affine transformation parameters between source image and each destination images. But these groups of parameter should be correlated, for it is from the same patch of terrain. This defect can be overcome by redefining the multiple images matching as: there is a plane surface or high order surface on the ground; it can be regarded as a mirror which reflects the pixel of source image to destination image. If plane patch is concerned, the parameters to be solved are the Z coordinate of patch center and the direction angle of the normal vector. There are only 3 parameters altogether for geometric deformation. Another advantage is the normal vector of the plane patch can be estimated.As we have got line segments and DSM, line matching can be processed by searchingthe match candidates in elevation range estimated from DSM, and 3D line segments can also be generated with forward intersection. 3D lines and 2D lines make a big network. The probability relaxation can be used to determine the correct match. The adopted compatibility measure between lines is the distance.From matched 3D lines, plane surfaces can be detected and fitted with least square. With semi-plane parameter space, best plane is detected in the rotation angle space. And plane is fitted with least square adjustment.A complete building model needs topographic relation to determine the adjacent plane to do intersection of planes. The results are point or line. At last, a combined least square adjustment can be adopted considering the image information, model condition, etc.