A Study Of3D Reconstruction Methods Based On Image Feature Points

Reconstructingthethree-dimensional(3D)structureofascenefromtwo-dimensional(2D) images is a fundamental problem in computer vision. After decades’ of extensivestudy of image-based3D reconstruction, this topic remains quite active as evidenced bycontinued rapid progress being made in the last decade. The task of image-based3D re-construction is the reverse process of image capturing, which corresponds to estimatingall (or some) camera parameters and3D locations of the scene points from their2D ob-servations. This problem poses many challenges due to unavoidable noise and outliers inthe data.Thisthesisstudiestheproblemof3Dreconstructionwithoutpriorsofscenestructurebasedonimagefeaturepoints,concentratingonfourtypicaltopicsrangingfromtwo-viewto multi-view scenarios and from land-based system to underwater environment. Aimingat improving the robustness, accuracy and efficiency of3D reconstruction, this thesismakes the following original contributions:The problem of3D reconstruction from two uncalibrated images is recast as a ro-bust single constrained optimization problem, which can be efficiently solved bya new hybrid optimization framework based on the modified ε Constrained Adap-tive Differential Evolution (εADE), within which both stability and convergencerate have been significantly improved by incorporating novel geometrically mean-ingful evolutionary operations. The above constrained optimization formulation isable to handle noise and outliers in image observations properly and to avoid geo-metric ambiguity in the reconstruction. Moreover, the above method considerablyoutperforms existing algorithms in terms of the accuracy of3D reconstruction.A new robust hybrid optimization framework for multi-view L2triangulation basedon optimal inlier selection and3D structure refinement is developed. In order to es-timate the scale of noise in image measurements, a new residual-consensus schemewithinwhichtheuncertaintyofepipolartransferisanalyticallycharacterizedbyde-riving its closed-form covariance is proposed. As for3D structure refinement, twonovel error bounding algorithms are proposed to significantly reduce the searchspace. Compared with state-of-the-art triangulation methods, the proposed frame-workisabletoobtainreconstructionresultswithmore3Dpointsofhigheraccuracy.The computational efficiency is also noticeably improved as validated by experi- mental results on large scale datasets.For an underwater imaging system, a refractive interface is introduced when a cam-era looks into the water-based environment, resulting in object space distortion inimage due to refraction. This thesis proposes a novel method for two-view under-water3D reconstruction based on refractive camera model, dealing with a generalunderwater imaging setup using two cameras, of which each camera is placed in aseparate waterproof housing with a flat window. To the best of our knowledge, theproposedframeworkisthefirstonecapableofperforminghighlyaccuratetwo-viewunderwater3D reconstruction without using any calibration object.This thesis for the first time presents theoretical analysis and also systematicallyquantitative evaluation of the influence of refraction on multi-view underwater3Dreconstruction based on perspective camera model. The results reveal a rather sur-prising and useful yet overlooked fact that the traditional perspective camera modelwithlensradialdistortioncorrectionandfocallengthadjustmentcancompensateforrefraction distortion, as long as system parameters satisfy some loose requirements.This research not only justifies the use of perspective camera model in multi-viewunderwater3D reconstruction, but also provides theoretical and practical supportfor the design of underwater3D imaging system.To summarize, thisthesisdevelopsanumberofnewapproachestoimprovethequal-ity of image-based3D reconstruction. The proposed methods have been evaluated withina complete3D reconstruction pipeline and their superior performance is verified by ex-tensive experiments and comparison with related state-of-the-art methods.