Study On The Key Matching Algorithms For Oblique Stereo Images And Its Application

Large oblique stereo images have been playing a key role in realistic3Dreconstruction of digital city due to the advantages of stable geometric imaging model,big coverage and abundant texture information. However, the perspective of sensorhas changed notablely in the process of acquiring multi-view images, result in theproblem of severe geometric distortions, radiometric distortions and correspondingregion shadow, which hinders the work of automatic detect conjugate points bycomputer. Therefore, there are significances to study the reliable matching algorithmfor this type of images with the aspects of improving the efficiency of realistic3Dreconstruction and promoting the automatic level of digital photogrammetry.Based on these above, research in this paper would mainly focus on fundamentaltheories for extracting and matching of affine invariant features, complementaryregion matching method with multi-type regions, Least Square Matching (LSM)algorithm with high accuracy and robustness, quasi-dense matching strategy based onperspective invariant neighborhood transform, multi-view subdivision matchingtechnology based on complex3D scenes, matching method for corner points ofbuildings under the constraint of epipolar driving and the application in3Dreconstruction based on above algorithms and strategies. The major achievements areas follows:(1) In order to evaluate the integrating performance of existing affine invariantfeatures, we proposed a novel Complementary Metric (CM), based on which thequantitative assessment results indicate that MSERs and Harris&Hessian Affine havebetter complementary performance; and in the further study, an automatic andadaptive matching method is implemented from coarse to fine based oncomplementary features, experiments on real oblique stereo images show that it canincrease the number of matches with uniform distribution.(2) In order to improve the robustness of region matching, the dominant gradientorientation of feature neighborhood is replaced by accurate epipolar ray orientation,and feature location is optimized by Least Square Matching (LSM), the initialiterating matrix of LSM can be obtained based on inherent geometric information ofcorrespondent affine regions, test results show subpixel precision is gained; based onLSM of affine invariant features, an novel quasi-dense matching method withperspective invariant neighborhood transform is put forward, the experimentsdemonstrate that the proposed algorithm has significant superiority in terms of accuracy and efficiency.(3) An precise matching algorithm through subdividing affine approximateplanes is proposed for oblique images, characterized by severe projective distortionsand complexity of scenes, this method is valid for achieving matches located indifferent depth of fields; and we also propose an matching method for corner points ofbuildings based on epipolar driving and self-adapted window, which is a feasiblesolution to obtain corresponding points in regions of disparity discontinuities.(4) An evaluation system of realistic3D reconstruction is established. We havedeveloped a set of turning points technology with automation and hierarchy based onintegrating aforementioned reliable matching methods, experimental results prove thatthe improved technology can provide sufficient correspondents for automaticaerotriangulation of multi-strip oblique images, finally, the method and illustration ofautomatic modeling with oblique images are displayed.