Research On Stereo Matching And Three-dimensional Reconstruction Of Urban Remote Sensing Image Pairs

With the rapid development of human society, more and more requirements of applications related to image have transited from traditional 2D plane images to 3D stereo images. Meanwhile, the reconstruction of remote sensing image attracts attentions all around the world due to the development of remote sensing technology and necessaries of relevant applications of earth observation. As a core technology in 3D reconstruction system, the results of stereo matching technology influence the accuracy of reconstruction results. Moreover, using amendments targets to image contents after primary reconstruction, deficiencies of matching results can be compensated and more subtle reconstruction results can be obtained. the proper amendments can also affect the final reconstruction results in some degree. Focus on the troubles and bottlenecks in traditional stereo matching and 3D reconstruction, this thesis makes further research on these aspects.At first, this thesis introduces the development histories of stereo matching and 3D reconstruction. During the introduction of stereo matching approaches, the thesis summaries the characteristics of various kinds of matching approaches according to the differences of matching units. Additionally, the thesis also analyzes the shortcomings of traditional stereo matching approaches when it is used to match urban remote sensing images. During the introduction of 3D reconstruction technology, the necessity and significance of this project are emphasized by comparing the performance and applicability between 3D reconstr uction approach of binocular vision system and other approaches.Then, the thesis researches the necessary preprocessing methods of urban remote sensing image pair matching. In order to shrink the searching space and reduce the matching time cost during the stereo matching process, the thesis researches related theories of epipolar line in photogrammetry, and then narrow down the searching space from 2D space to 1D space by using epipolar constraint. For urban area that contains massive buildings, edge information of buildings is the most abundant features, so in preprocessing part of the thesis we proposes a method to extract building edges. This method includes three steps, which are edge-preserved image filtering, line segment extraction and line segment amendment. The method is proved to be able to extract more realistic building edge line segments than other common methods from urban remote sensing images through experiments, and it lays a foundation for the subsequent stereo matching.Before proposing the new stereo matching approach, the thesis studies the traditional stereo matching methods and analyzes the causes of their shortcomings when they are used for urban remote sensing matching. For the purpose of obtaining better stereo matching results, the edge line segments extracted in the preprocessing are matched by line matching approach firstly, and utilizing the matching results as a constraint to further narrow down the searching range. Then, through comparing the differences between multiple windows template and single window template, the advantages of multiple windows template are reflected. By combining the multiple windows template and adapted window method, a novel edge-protected stereo matching approach of urban remote sensing image pair is proposed, which improves the matching accuracy of pixels in building areas.Finally, the least square method of RFM based ground elevation resolving is researched for the purpose of achieving elaborate urban area 3D reconstruction. After verifying accuracy of the resolving method, it is found that reconstruction result have great errors in shadow and around building edge areas when comparing the reconstruction results of real scenes with Li DAR data of the same district. For errors in shadow areas, the thesis extracts shadows first, and then a neighbor minimum with ratio threshold method is adapted to fit the height vales. For errors around building edge areas, a novel building rooftop contour extraction based on classification and parallelogram segmentation is proposed when building edge line segments obtained. The proposed method can extract building rooftop contour quickly and correctly, and the extraction results are utilized to amend the final reconstruction result to make the average accuracy of 3D reconstruction over 90%.