Feature Points Matching And 3D Reconstruction Of Object

In recent years, along with the development of computer and image processingtechnology, applications based on these develop very fast, such as image featurematching, face recognition, iris identification, gesture recognition, AOI technology,target tracking and so on. At the same time, computer vision has also achieved greatprogress. Computer vision is the simulation of human vision. The real world isthree-dimensional but the image obtained by camera or human beings'monocular visionis two-dimensional. Human could perceive three-dimensional information frombinocular, so computer vision is focused on simulating this ability to build a 3D modelof a scene based on two or more images or video sequences. That is what we call 3Dreconstruction. Computer stereo vision emphasizes the relationship between 2D imageand 3D scene.The frame and three main steps of 3D reconstruction are described as below:i) Compute the fundamental matrix based on matching points;ii) Compute the projection matrix based on fundamental matrix;iii) Compute the 3D coordinate of points based on matching points and projectionmatrix.From the above conclusions, we realize that matching points is the premise of 3Dreconstruction, and the same to many other computer vision applications. It is essentialand prior to other steps.First, a simple survey about feature points matching,camera calibration and 3Dreconstruction is made. After a deep research on SIFT algorithm, we analyse theadvantages and deficiencies of SIFT. In order to complement the deficiencies of SIFT,we propose the corner matching algorithm based on SIFT matching points. Twoimprovements are achieved. First, multi-scale corner detection is realized based onGaussian Pyramid generated in SIFT algorithm. Second, the candidate search range isrestricted from global to local. High matching accuracy can be obtained by just usingsimple feature vector. We focus on the research with regard to matching, points cameracalibration and computation of projection matrix. We estimate the fundamental matrixusing SIFT points. Based on camera parameters calculated through Zhang's method, wetransform the fundamental matrix into essence matrix, and then obtain the projectionmatrix ground on matrix decomposition theory. Finally, we reconstruct the matchingcorners; accomplish the 3D reconstruction of an object.Following the reconstruction frame and steps, we focus on the research with regardto feature point matching, camera calibration and computation of projection matrix. Atlast, we complete the reconstruction task. The main investigations are described asbelow.1,We make a simple survey about image matching. After a thorough research about image matching algorithms, we propose the corner matching algorithmwhich is based on SIFT. The SIFT matching points are derived at first, then wesearch the corners in the limited range which is determined by SIFT points asthe center. We test the algorithm in many conditions such as Gaussian noiseswith differentσ, different kinds of noise (Salt, Poisson and Speckle), imageswith different scale, image with different rotation angle. All prove the goodperformance.2,We estimate the fundamental matrix using normalized eight points algorithmbased on SIFT matching points. Fundamental matrix is transformed intoEssential matrix with the help of camera calibration. We get the projectionmatrix through the decomposition of Essential matrix. Then we check the rightone through four candinate projection matrix. Finally, the 3D reconstruction ofan object is achieved by reconstructing corner points using the right projectionmatrix.