The Research And Application Of Oblique Images Matching Methods Based On Invariant Features

Oblique photography can both get the integrity images and the facade textures ofbuildings. It is the mainstream trend of3D modeling of urban landscape. Due to the hugegeometric distortion, the existence of chromatism, the large amount of image data, these allmake the robust and rapid matching for oblique images a difficult problem.In this dissertation, in order to deal with the questions about scale, rotation, displacement,perspective deformation which can be simplified as local affine deformation of obliqueimages, we research on the matching methods for oblique images based on invariant featuresand its applications. The main research contents and innovation points in this dissertation arelisted as follows:(1) The introduction of tilt photogrammetry, including the emergence and developmentof tilt photogrammetry, the introduction of oblique multi-camera systems, the characteristicsand functions of oblique images. The thinking and discussion about the matching of obliqueimages, including the research status at home and abroad, the difficulties andcountermeasures of oblique images matching. The summary of image local invariant featuresthat gives a detail introduction about some of detection algorithms, descriptor algorithms andmatching strategies in the image matching processes based on local invariant features.(2) ASIFT is a good matching algorithm for oblique images, but it is low efficiency. Wemake full use of the prior knowledge and propose a matching algorithm for oblique images based on affine invariant features, which named AIF. In the general case, the matching effectof AIF algorithm is slightly better than ASIFT algorithm, and its matching efficiency hasbeen improved more than100times.(3) The distribution of the corresponding points by AIF algorithm is not uniform. Weuse perspective transform to take place of local affine transform and propose a new matchingalgorithm based on perspective invariant features, we called it “PIF”. Not only thedistribution, quantity and accuracy of the corresponding points by PIF algorithm, but also thematching efficiency of PIF algorithm are both superior to AIF and ASIFT algorithms.(4) During the matching process, we use the fundamental matrix and homography matrixto calculate the potential area of the corresponding point, and pick out the Harriscorner-points in this area as candidate points, then use the Nearest Neighbor DistanceRatio(NNDR) and Normalized Cross Correlation (NCC) measure constraints to get thematches. This strategy makes the distribution of matching points is uniform and the quantityis large, at the same time, it also improves the matching efficiency and reduces themismatching rate.(5) In order to improve computational efficiency of bundle adjustment for obliqueimages, we propose a bundle adjustment model for oblique images which take the relativeattitude parameters of cameras into account.(6) We apply the PIF algorithm to the full-automatic aerotriangulation, densepoint-cloud matching and3D model generation for oblique images. Experiments show thatthe precision of aerotriangulation can meet some application requirements, the idea ofautomatic3D reconstruction is feasible and the result of automatic3D reconstruction canmeet a certain degree of visual demand.