Pose Priors for Aerial Image Registration

RANSAC has become an essential tool for the robust estimation of inter-image geometries using point correspondences. Due to its randomized nature, RANSAC becomes inefficient for large numbers of incorrect matches. This paper presents four methods for utilizing prior geometric information when estimating the homography which register a pair of near-nadir aerial images. In this case the initial geometric estimate is derived using strict assumptions about the camera and scene geometry and photometric information present in each image, however, the techniques described could be applied to other sources of pose prior information. The algorithms described allow for the successful estimation of the inter-image homography in a severely constrained number of iterations, even in the presence of low numbers of inliers.