Research On Some Key Issues For3D Reconstruction Using Multi-view Images

Multi-view3D reconstruction is one of approaches for3D modeling in computergraphics and computer vision, which aims to abstract the scene and scenery objectsfrom an image or a series of images. It is actually critical for improving the sense ofphotorealistic3D modeling and building the3D model for real-time scenery. Currently,relevant research achievements are used in the worlds of digital medicine image, digitalcity, autonomous robot, and digital entertainment. However, the technology ofmulti-view3D reconstruction still has certain problems that should be solved inreal-world applications. This dissertation explored the problems related with themulti-view3D reconstruction, and proposed several effective solutions to the relevantproblems. This dissertation focuses on extracting the feature vectors problems and thematching problems of images, counting the limited geometrical relationship, estimatingthe parameters and the dense surface robustly, as well as rebuilding the scene throughstructure from motion. This dissertation makes the following contributions:(1) An approach called as new DMD (Differential Morphological Decomposition)is proposed to detect and describe local invariant features of the image effectively, whenextracting robust feature points in a series of images with complicated scenes, as well asimages have some complicated features, such as multiple dimensions, fuzzytransformation and both of them. Being aware of these drawbacks of the existingmulti-scale feature-detecting algorithms in the scale space theory, this dissertation firstproposed a pyramid scale space, which is adopted by DMD. And then applied a Harrisoperator on scale images in the pyramid scale space, so that feature points can bedivided into deferent groups, and every group of points just needs to describe a localstructure of the image. Finally, a PCA-SIFT descriptor is leveraged for featuring andmatching a feature point, which is selected by a LoG value in a scale domain. Theexperiments with real-world and artificial data sets have confirmed that the proposedapproach achieved better results in detecting and describing local invariant featurepoints in the cases when scale fuzzy transformation and luminance transformationoccurred. (2) Afast and accurate RANSAC (Random Sample Consensus) is proposed, whichis based on SPRT (Sequential Probability Ratio Test) and local optimization technique.At first, in order to improve velocity of RANSAC, this dissertation employed SPRT andcertain preview evaluation methods to optimize the procedure of model verification. Tothis end, the proposed mechanism randomly selected a few of sample data from sourcedata sets and then made statistical verification for these data, this process is also calledas preview evaluation. Therefore, the final verification will be conducted on the wholedata set when the preview evaluation has passed; otherwise, no tests will be done. Thesecond motivation is to improve the accuracy of RANSAC. After the previewevaluation, inliers (the data that is fit the hypothesis in the supposed model) can beexplored; then, the explored potential inliers sets can be used in a local optimizationRANSAC with the selected optimal model. The experimental results have disclosed thatthe proposed approach can gain lots of benefits from the theory of RANSAC with thenewly proposed approaches.(3) Anew algorithm is designed for conjugate gradient bundle adjustment which isbased on the block preconditioned and EPIs (Embedded Point Iterations). Whilereconstructing a large scale of scene, the method of bundle adjustment is a well-knownbottleneck during computation. In this newly conjugate gradient method, the maincomputational tasks include a simple matrix and vector multiplication with the jacobian.Furthermore, this newly presented algorithm has an alternative method, which iteratesinteriorly with a conjugate gradient algorithm. In this dissertation, for enhancingefficiency of the proposed bundle adjustment algorithm by using conjugate gradient, thealgorithm first decreased the cost of per iteration to almost a half by completelyutilizing a property of the least square method, because the improved method employsan easy preprocess system of QR factorization based on block preprocess. Besides, inorder to accelerate processing speed, by using the EPIs method, this algorithmembedded iteration points into every camera correction phase, so that the cost of everycamera correction step can be reduced. The experimental results show that whilerebuilding a large-scale scene, this newly proposed algorithm outperforms otherapproaches.(4) This dissertation presented a new method that computes a dense3D point cloudfrom depth maps. In order to realize surface reconstruction from sparse3D points, this dissertation presented an improved PMVS (Patch-based Multi-view Stereo) quasi-densealgorithm based on geometrical constraint and self-adaption. At first, this algorithmassumes that the information about location and orientation of the camera is available,and then it computes depth maps by solving a global energy minimization problem in animage space, so that each pixel in a depth map can be projected back into a common3Dspace to yield an extremely dense point cloud, which contains of millions of points.Eventually, this point cloud can reconstruct a surface mesh by using Poissonreconstruction. Secondly, aiming at the issue of3D reconstruction from a series ofuncorrected images, the mechanism of SFM (Structure From Motion) can onlyreconstruct sparse3D points. It is well known that although these points are enough fortracing the camera’s position, it is not enough for reconstructing the scene or objectswith fairy accuracy and authenticity. This newly presented algorithm adds geometricalspace constraints and adaptive expanding algorithm into the procedure of patchexpanding in PMVS algorithm. Therefore, it can generate robust and accurate geometryestimation, and obtain a highly dense3D point cloud with fewer images. As a result,this algorithm obtains realistic3D models of the scene and objects by using surfacereconstruction approach. The experiments have shown that this improved algorithm isable to rebuild accurate surfaces as far as possible and improve authenticity of the3Dmodel.