Efficient Registration of Aerial Image Sequences Without Camera Priors

摘要

We present an efficient approach for finding homographies between sequences of aerial images. We propose a two-step approach: a) initially solving for image-plane rotation and scale parameters without using correspondence (under affine assumption), and b) using these parameters to constrain the full ho-mography search, and c) extending the results to full perspective projection. No flight meta-data, camera priors, or any other user defined information is used for the task. Based on the perspective parameters estimated, the aerial images are stitched with the best matching image based on a probabilistic model, to compose a high resolution aerial image mosaic. While retaining the improved asymptotic worst-case complexity of [6], we demonstrate significant performance improvements in practice.