Abstract: Global motion estimation is a critical problem in aerial video mosaicing. A more simple and robust linear similarity transformation model with only 4 degrees of freedom is proposed to satisfy the relationship between two consecutive frames. And the transformation between two non-contiguous frames can be obtained by frame-to-frame recursion. Feature-matching-based approach is adopted for image registration. Once scale invariant feature transform (SIFT) features extracted, they are matched by a novel matching scheme—bidirectional nearest neighbor distance ratio. Finally, the motion parameters can be estimated by random sample consensus (RANSAC) algorithm. The experimental results show the high accuracy of the global motion model for aerial video mosaicing and the good performance of the proposed matching scheme.

Key words: image processing, global motion estimation, scale invariant feature transform, feature matching, image registration, parameter estimation