Research And Implementation Of Multi-view Dynamic Three-dimensional Reconstruction Techniques

Three-dimensional reconstruction technique has been one of the computer visionfield research hotspot in recent years, with the rapid development of a variety of3Dindustry, people’s vision gradually shift from2D products to3D products, whichpromotes the development of three-dimensional reconstruction techniques. It is moreimportant to achieve the scene reconstruction which is large-scale and complex, andhas a broad application prospects, such as the television and entertainment, modernindustrial and manufacturing, health care, defense and military fields.This paper focuses on the key technology research of multi-view dynamicthree-dimensional reconstruction. Our acquisition system is a diameter of10metersand a height of7m dome-type architecture, the built-in camera is composed of20circular camera array. In this paper, an annular array camera, designed a workablecamera calibration solutions. For visual hull reconstruction, the paper based on theepipolar geometry and caching technology, presents an efficient reconstructionmethod. Meanwhile, this paper presents a multi-view dynamic three-dimensionalexact reconstruction algorithm, and different algorithms with the general MVS, whichoptimize the use of local area extraction point cloud and the realization of a dynamicthree-dimensional reconstruction. The major contributions of this paper are listed asfollowed:(1) Chapter2designed a ring-shaped camera array calibration solutions. First theacquisition mode of calibration board image in different ways with the general cameracalibration, this paper is to keep the camera position unchanged, shooting differentposes in different locations calibration plate. And different calculation methods, thispaper is the first of a sequence of images acquired grouped within an outer set ofcomputing camera parameters and internal control, and finally between the differentgroups to obtain the calibration results by adjacent outer parameter conversion.(2) Chapter3presents a highly efficient visual hull reconstruction algorithm.Compared with the general visual hull reconstruction algorithm, the proposedalgorithm is computationally optimized to do both, which greatly improves theefficiency of the algorithm. On the one hand, the use of the intersection between theline and polygon instead of the intersection between the polygon. On the other hand,the use of caching mechanism to avoid duplication of traversing the epipolar line, saving a lot of computing time. In order to obtain the desired contour information ofthe algorithm, this chapter implements a foreground-background segmentation. Andestimating the point cloud’s normal with principal component analysis. Finally, weobtain a three-dimensional point cloud model with normal.(3) Chapter4presents a multi-view dynamic three-dimensional exactreconstruction algorithm. The algorithm mainly research the robustness of the pointcloud matching algorithm, the edge points, the detection of hidden point and theintegration, denoising multi-view cloud, Poisson reconstruction based on thestructured surface and other issues, we use the point cloud which reconstructed byvisual hull based on the contour as the input data, after the extraction, integration andtriangularization of the point cloud, we achieve the reconstruction of the gridthree-dimensional model, and achieve a dynamic three-dimensional reconstruction.(4) Chapter5presents a texture mapping method for the result of the MVSalgorithm in this paper, and realize the texture draw of the3D model by the using ofOpenGL, achieving a very realistic color three-dimensional models.