Research On The Automatic 3D Surface Reconstruction Of High-Resolution Urban Satellite Images

Automatic 3D reconstruction of large-scale urban scenes based on satellite images has been a hot research topic in the intersection of photogrammetry and computer vision.As satellite images have the inherent advantage of global coverage,3D reconstruction based on satellite images plays an irreplaceable role in global 3D information perception,national 3D geographic security and urban construction.At present,the research on 3D reconstruction of urban satellite images is still immature,and many algorithms are directly borrowed from the3 D reconstruction of aerial images without adapting to the characteristics of satellite images.In particular,there are many problems to be solved in stereo matching based on two-view satellite images,3D reconstruction based on two-view satellite images and 3D surface reconstruction based on multi-view satellite images.In this thesis,the research objective is to achieve the 3D surface model of high-resolution urban satellite images,and to carry out the research on 3D reconstruction based on satellite images from two views to multiple views.The main key problems to be solved in this task include the super pixel segmentation of satellite images,the stereo matching of urban satellite images and the multi-view DSM fusion,etc.The main research works of this thesis is as follows:Firstly,to solve the interference caused by depth discontinuity regions,weak texture regions and occlusion regions in urban satellite images to the 3D reconstruction of two-view satellite images,a superpixel algorithm via region fusion with boundary constraint has been proposed for satellite images in this thesis.In this way,accurate region and edge constraints can be provided for two-view satellite image 3D reconstruction by superpixels.The proposed superpixel algorithm includes two steps which are initial segmentation and region fusion respectively.In initial segmentation,the sufficient edge information and region information can be obtained through the proposed growth-based edge closure algorithm.In region fusion,similar region fusion and true edge preservation are achieved by means of inter-regional similarity message propagation,and the regular shape of super pixels in urban satellite images is also recovered.Experiments show that the proposed algorithm can adapt to the complex geographic environment in urban satellite images and obtain superpixels with regular shapes and accurate edges,which can provide accurate superpixel results for two-view satellite image3 D reconstruction.Then,to overcome the interferences caused by the depth discontinuities,weakly textured areas,and occluded areas,a double propagation stereo matching algorithm is proposed for 3D reconstruction based on two-view satellite images in this thesis.The algorithm first calculates the initial matching cost based on three matching costs,iteratively optimizes the initial matching cost using the proposed double propagation optimization model under the constraints of superpixel edges and geometric model,and obtains the disparity maps by minimizing the energy function.Experiments show that the proposed algorithm is not only able to obtain accurate parallax values in depth discontinuity regions and weak texture regions,but also to calculate digital surface models and 3D point clouds that match the actual terrain based on their disparity maps.The proposed algorithm can also provide accurate digital surface models for 3D surface reconstruction of multi-view urban satellite images.Finally,to solve the problems of fuzzy elevation edge delineation,severe elevation holes and loss of detailed surface structure,a multi-view digital surface model fusion algorithm based on neighboring point correlation is proposed for 3D surface reconstruction of multi-view urban satellite images in this thesis.NPC adopts a coarse-to-fine strategy that firstly obtains the localized correlation among edge points through superpixel segmentation based on 2D-3D joint data.After that,geometric correlations among discrete 3D points are established by surface model clustering.Such that,the spatial and geometric correlations are used to locate the edges of buildings with severe elevation drop-offs and to recover fine surface structures in weakly textured areas.Experiments show that both the elevation divisions with sharp elevation drop can be enhanced under the constraint of superpixel edges,and the fine surface structures hidden in weakly textured regions are able to be uncovered guided by surface models.Eventually it is possible to reconstruct a 3D mesh model of the urban surface based on the digital surface model of the urban scenes.