Parameter Estimation And Shape Reconstruction Based On Multiple Views

Computer vision is a comprehensive discipline whose researches relate to image processing, image comprehension, pattern recognition, computer graphics, signal processing, mathematics and biological physics, etc. The research object of computer vision is to make computers have the ability of understanding 3D environmental information from 2D images or views. Because the research achievements can be applied directly on robot localization and navigation, precise industry measurement, object recognition, visual reality, military affairs and many other fields, the researches on computer vision problems have become one of the most popular research subject in the world.This thesis deals with many popular problems in the field of computer vision. The major researches are hierarchical reconstruction based on multiple views and uncalibrated P5P problems. In the first part, depending on three or more images, the main research work are listed as follows: (l)using SVD decomposition to realize projective reconstruction; (2)realizing camera self-calibration by solving Kruppa's equation; (S)recovering Euclidean reconstruction from projective reconstruction. Depending on only two images, the main researches are: (l)making out infinite plane homography matrix by using scene structure information, then recovering affine reconstruction from projective reconstruction; (2)making out the absolute conic images by using scene structure information, and then recovering Euclidean reconstruction from projective reconstruction. The second part of this thesis mainly research on solving the uncalibrated P5P problem when the camera intrinsic parameter is unknown and it is changeable in the camera motion.The main research achievements are following:1. From camera model, we deduce the relation between perspective model and affine model, discuss the properties of projective depths, introduce the general frame of projective reconstruction based on SVD, analyze and realize the estimation of projective depths based on fundamental matrix and epipolar points. Taking the measurement matrix rank 4 as the constraint, we propose two methods to estimate projective depths iteratively: (l)the algorithm of estimating projective depths based on conjugate gradient method; (2)the algorithm of estimating projective depths based on genetic algorithms. After obtaining correct projective depths, we decompose the measurement matrix into camera motion in projective space and projective reconstruction by SVD. The experiments show that the method proposed by us is more robust, comparing to the method of estimating projective depths based on fundamental matrix and epipolar points.2. We discuss the traditional camera self-calibration methods based on Kruppa's equations and propose a new method for solving Kruppa's equations-step method.IllFirst we use conjugate gradient method to estimate the unknown scale factors in Kruppa's equations, then use the scale factors to solve Kruppa's equations linearly and calibrate the camera intrinsic parameters. We propose an algorithm of recovering Euclidean reconstruction from projective reconstruction if the camera intrinsic parameters are known. First solving a non-singular matrix which satisfies Euclidean reconstruction conditions and then we convert the projective reconstruction to Euclidean reconstruction by the matrix. The experiment results show that our algorithm is feasible.3. We analyze the essence of affine reconstruction and prove the sufficient conditions that a reversible matrix can be an infinite plane homography matrix and we can not uniquely decide an infinite plane homography matrix from fundamental matrix. We systemically discuss how to uniquely decide an infinite plane homography matrix by using the structure information in scene and how to evaluate a homography matrix which convert affine reconstruction to Euclidean reconstruction by solving absolute conic images. We give three constraints of absolute conic images and use these constraints to evaluate absolute conic images and th...