Study And Implementation Of3-D Reconstruction Based On Image Sequence

3-D reconstruction is basic component both in computer vision andphotogrammetry, and many methods could achieve this goal. But, compared withclassical reconstruction ways relying on modeling software or other expensive hardwareequipment like3-D scanner, the method based on image sequence has many advantagessuch as simple operation, flexible application and low cost. What’s more, a vivid andobjective result also can be obtained.3-D point cloud reconstruction is regarded as concrete target based on imagesequence for3-D scene in this thesis, and many key technologies related to the wholeprocess are researched. Analyzing methods of3-D reconstruction based on imagesequence, the whole process is divided into four main parts: feature points detecting andmatching, robust calculation of fundamental matrix, camera calibration and3-Dreconstruction strategy. Then algorithms about that four parts are studied, analysis,improved and implemented in this paper. Meanwhile intrinsic relationship betweencomputer vision and photogrammetry is also found on some main content. Main worksof this thesis are listed as follows:(1) Feature points detecting and initial matching and camera calibrating are achievedFeature points detecting and initial matching is accomplished using SURFalgorithm, with characteristic of scale invariance and resistance to affine transformation.Camera calibration is accomplished by Zhang Zhengyou method.(2) Proposed a robust fundamental matrix calculation algorithm based on Gaussiankernel fuzzy clusteringConsidering distribution of matching points’ residuals, residual is selected asclustering feature. Then residuals are mapped into a high-dimensional feature space, forbuilding linear dividable features. As a result, matching points could be divided intoinliers and outliers via fuzzy clustering method. Then a separability judgment criterionbased on Chernoff error bounds of Bayesian classifier is constructed. Finally, stableinlier set is obtained and precise fundamental matrix is computed.(3) Proposed a3-D reconstruction strategy of image sequence based on three-viewConsidering matching point can’t be judged absolutely by its residual just rely onepipolar geometry, optimum inlier set couldn’t be selected all the time, thenfundamental matrix also is influenced. Three-view constraint has stricter limitation. Soit is applied to select more accurate inlier set and acquire precise result of fundamental matrix. Then3-D point cloud structure of three-view is rebuilt. Stable and rigid body aslocal structure is, a novel reconstruction strategy is proposed:3-D point cloudcoordinate of whole image sequence cloud be unified just by transforming betweenadjacent three-view.(4) Compared and analyzed key technologies from aspects of computer vision andphotogrammetryIntrinsic relationship between computer vision and photogrammetry is built byderivation of mathematic formula and interpretation of physical meaning for someconcepts. Consequently, conclusion could be drawn safely that computer vision hasclose relationship and high similarity with photogrammetry about3-D reconstructionbased on image sequence. So they could improve themselves by learning from eachother.