Rough Landform Guided Semiglobal Optimization For Dense Image Matching And Terrain Surface Reconstruction

Since the implementation of the national special project "high resolution earth observation system",China has made a series of achievements in remote sensing science.However,in quite a long period of time we are devoted to acquire remotely sensed data without considering the needs analysis and application research.Thus we failed to process the massive remote sensing data efficiently.Therefore,how to process the massive high resolution remote sensing data accurately and mine spatial information intelligently has become the key problems of realizing the economic and social value of the earth observation system.Under this background,the national key basic research program(973)projects and the National Natural Science Foundation funded a lot of subjects which is aimed at automatic extraction of 3D geometric information from high resolution remote sensing data,including 973 sub project "Natural Terrain and Artificial 3D Reconstruction from High Resolution Remote Rensing Images",Natural Science Fund Project "Digital Photogrammetry and Remote Sensing" and "Theory and Methods on Single Stereo Model based Multi-view Imagery Dense Matching".Based on the above three topics,the main purpose of this paper is to study the theory and method of automatic extraction of 3D surface model from high-resolution remote sensing images.The main work includes the following four aspects:(1)The semi global optimization method is studied.Many image processing problems can be transformed into Markov Random Field(MRF)labeling problem,which can be solved via minimization of energy function defined on MRF.Semi global matching(SGM),as an efficient method for solving energy function,has achieved great success in stereo matching.But as an optimization method,the nature of semi global optimization has not been systematically analyzed and clarified.In this thesis,the relationships between semi global optimization and MRF,dynamic programming and belief propagation are explored.Then 7 factors of semi global optimization are analyzed.Reasonable consideration of these factors is necessary to solve specific problems using the semi global optimization.(2)Object space based DSM generation by multi view dense matching is studied.The existing DSM generation process usually implements dense matching using two or more views,which is involved in a complex workflow with high computation redundancy.In this thesis,a novel semi-global vertical line locus method(SGVLL)is proposed to generate a DSM directly in the object space.The method has two contributions:1)a novel workflow for DSM generation is proposed.Different from the traditional DSM based on image space,the proposed method works in the object space,and thus it is much simple.Meanwhile,the results of the proposed method are as reliable as the image space based methods;2)the standard VLL is majorly improved.By designing robust adaptive elevation step calculation and occlusion detection to improve the robustness of object cost calculation and introducing the rough landform to guide the semi global optimization,the proposed method achieve state-of-the-art results.(3)The method of DSM refinement is studied.Due to over-smooth or coarse surface and loud region of cloud or water areas,DSM needs to be refined.Two contributions have been made in this thesis:1)Based on the DSM reliability,the large cloud and water areas of DSM is detected,and then repaired seamlessly using Poisson fusion method;2)semi global refinement(SGR)method and a mean shift fusion of depth maps are proposed to refine the original DSM.(4)Automatic filtering of the middle and low resolution DSM is studied.Although a lot of cloud filtering methods have been proposed to extract DTM,most of them are designed for high density point clouds,and are not suitable for middle or low resolution DSM.In the current situation,there are still a large number of middle or low resolution DSM,from which DTM needs to be extracted.However,it is usually done by manual editing.Therefore it is necessary to study how to extract a DTM from middle or low resolution DSM in order to improve production efficiency.The key point of this topic lies in the difficulty to distinguish the topographical features on the hills and the objects on the flat terrain since the former need to be kept while the latter need to be removed.This thesis proposes a two-step semi-global filtering method(TSGF)which separates the two kinds of components,then extract a DTM from middle or low resolution DSM.The contribution of this method includes three aspects:1)a framework of "two steps filtering" method is established to realize automatic filtering of middle or low resolution DSM;2)the automatic acquisition of flat region mask is realized by using SRTM and semi global filtering;3)a segmentation constrained semi global filtering method is proposed to extract a DTM from middle or low resolution DSM.To sum up,all the proposed algorithms including DSM generation,refinement and filtering adopt the idea of rough lanform guidance and semi global optimization.In order to verify the effectiveness of these algorithms,we conducted two aspects of experiments.First,independent experiments are designed to evaluate the key techniques involved in the proposed algorithms.Second,comprehensive experiments are designed to evaluate the final results of proposed algorithms using multiple sets of high resolution remote sensing imageries with various platforms,resolution and land cover types.Qualitative and quantitative comparison are made with corresponding state-of-the-art algorithms.All experiments have shown the effectiveness of the proposed algorithms,which also suggests that the theory and method of the rough landform guided semi-global optimization for dense image matching and terrain model reconstruction are correct and effective.