Automatic registration of aerial image series using geometric invariance

Abstract

In this paper, a strong emphasis is laid on the study of automatic alignment and registration technology for the purpose of stitching serial images and getting a larger, seamless and high-resolution one. Aerial image taken by robot flying flat has low quality and complex distortion, it is hard to find and evaluate transformation factors in the scene which appeared between the aeronautic consecutive images for mosaicking. So the registration turns on the crucial step of successful image mosaic. Based on the FBM and ABM algorithm as the initial matching, a matching technique of epipolar geometric theory is introduced as the final one in the interest of attaining more precise aggregate of matching points. In order to accurately and robustly estimate the F-matrix (fundamental matrix) which encapsulated the whole epipolar geometry, an improved SVD decomposition with weighted normalized F-matrix calculating method is proposed. By using geometric invariant, cross ratio to rectify the matching points, the rectification is realized automatically and the coarse error is reduced effectively.