Researches On Several Problems In Three-dimension Computer Vision

In this paper. we mainly discuss several prob1ems in the field of 3-D computervision, which are the algorithLms for fundamental matrix, conic correspondence, theestimation of projective depths and sphere orientation, etc. Our main work is asfollowing f(1) Robust algorithms for fundamental matrix. We firstly introduce severalpresentations of fundamental matrix, which are theoretical foundations of thealgorithms for fundamental matrix. As to two images, we give the algorithIns whichpreliminarily process data to inhibit noises. Then we give robust algorithms. In theend, based on the projective constraint of three images, we propose asix-point-RANSAC algorithm.(2) The a1gorithIn for conic correspondence. The existing algorithms for coniccorrespondence rely on a poIynomial condition. They are non-linear algorithms. Wegive a constraint for conic correspondence only from fundamental matrix, and thenwe design an aIgorithm fOr conic correspondence based on the constraint. ThealgorithIn is linear. Both simulated and realistic experiments verify our algorithm isright and feasible.(3) The estimation of projective depths. In the process of hierarchicalreconstruction, making out the prOjective depths is a key step. The existing algorithmsare effective to simulated data, but to realistic images, they can not effectivelyestimate the prOjective depths. Because genetic algorithIns are of advantages in theiriterative mechanism and evolutionaI mechanism, we apply them to design anaIgorithm to estimate the prOjective depths. The realistic experiment verifies thefeasibility of the algorithm.(4) Sphere orientation. Sphere orientation is to make out the 3-D coordinate of theIIsphere center and the sphere radius. Because every point on the surface of a spherehas the same geometric features, if there are no clear textures, we can not use featurematching to realize sphere orientation. We propose a new method in this papef, whichuses the cofltours of two sphere images to realize the orientation. The methodovercomes the defect of classical algorithIns, which can not realize the orientation ifthere are no clear textures on the surface of a sphere. Our method is simple and can bereaIized easily. The results of experiments also verify its effectiveness.