Research On3D Reconstruction Method For Stereo Vision

Recovering3D shape from2D images is a key research area in computer vision.For3D reconstruction of rotation stereo vision, the axis of rotation must bedetermined first. The space coordinates of a calibration object in a different position ismeasured first when the calibration object rotates on the rotating platform, then theaxis of rotation can be obtained by fitting Space circle. This is commonly usedmethod, but a lot of calculation in three-dimensional space is involved. For3Dreconstruction with uncalibrated image sequences, the commonly used method is socalled factorization algorithm. Before factorization, feature tracking must be obtained.Unfortunately, no perfect and truly general solution of feature tracking has been foundso far. In order to get the rotation axis and3D reconstruction with uncalibrated imagesequences, the main work and achievements of this paper is as follows：Firstly, research on determining the axis of rotation. We first determine the imageof discrete points on the rotation axis, corresponding discrete points in thethree-dimensional space can be obtained. By fitting these discrete points inthree-dimensional space, we can get the rotation axis. The method is different fromtraditional approaches. In the case of constant intrinsic parameters of camera, themethod of determining the rotation axis by fitting the quadratic curve has been deeplyresearched and we obtained Statistical error in different levels of noise. In the case ofconstant intrinsic parameters of camera in two locations，a method of determining therotation axis based on homography is proposed and the experiments showed that Thismethod is robust.Secondly, research on3D reconstruction with uncalibrated image sequences.This paper researched some elements of the three-dimensional reconstruction system.SIFT algorithm and eliminating the outliers are researched first. Then we researchedfactorization algorithm based on computing the projective depths from fundamentalmatrices and epipoles. In order to analyze the effectiveness of this method, wecomputed the angle of the parallel plane and the perpendicular plane. A method forprojective reconstruction Based on the initial reconstruction is proposed and the corresponding objective function was given too. For error accumulation problem inthe reconstruction method, three strategies are given. Analysis of the method is thesame to the above experiment. In order to get metric reconstruction automatically,self-calibration based on absolute dual quadric is researched at last.