Geometric Constraints In Image Sequence And Their Applications

This dissertation describes our investigation of the geometric invariance and geometric invariant over projective transformation group, with the emphasis on the estimation of geometric constraints in image sequence and their applications in computer vision. The geometric constraints in image sequence mainly include epipolar geometry ( bilinear constraint ) of 2 images and trilinear constraint of 3 images. The motivation of this dissertation is to obtain the estimation of the constraints with high accuracy and introduce their applications into the fields of computer vision and image processing. The main contribution of this dissertation includes 1. On researching the previous linear and nonlinear methods to estimate fundamental matrix, we present the 6-point synthetic method based on a new constraint; 2. We also solve the redundant problem between the matching points which definitely exists in the computation of fundamental matrix and is often neglected by the researchers. The development of these methods has significantly increased the accuracy and stability of the fundamental matrix obtained. Then a method to estimate the trilinear constraint from fundamental matrix is proposed. It has some advantages over the previous methods, for instance, simplicity in complementation and high accuracy in estimation. Based on the robust estimation of epipolar geometry and trilinear constraint, the dissertation gives some researches on their application in the field of computer, especially in the image matching, 3D reconstruction and view synthesis. Among these researches, an image dense matching method is developed using the epipolar geometry and DI triangulation. Take the advantage of the dense matching method, a view synthesis method is completed using pixel transferring technique. Based on geometric invariant theory and some image processing techniques, we have also developed a new view synthesis method from triangulation and quadrangulation. Finally, from many simulations and comparison with some classical and advanced methods in the relative fields, it is proved that the methods in the dissertation obviously improve the estimation accuracy and stability of the geometric relationship in image sequence. At the same time, it is shown that the geometric relationship in image sequence has extensively applied to the computer vision and other relative fields, and the applications will bring fruitful many good results.