Under the assumption of a non-translating camera, there exists a perspective transformation (homography) that globally aligns each pair of consecutive video frames. The registration module makes use of recent theoretical advances in sparse representation and compressive sampling. The sparse registration method is parameter-free and avoids explicit structure matching, by matching entire images or image patches. It assumes that outlier pixels (outliers to the global homography between consecutive frames) are sparse in an image. Consequently, the sparse registration method becomes equivalent to solving a sequence of L1 minimization problems, each of which can be solved using the Inexact Augmented Lagrangian Method (IALM). The sparse registration method avoids the instability of detecting and matching points, lines, or other primitive structures by matching entire images or image patches. No explicit correspondence between features is required. The matching process computes an optimal homography that maps one image into the other by assuming that outlier pixels are sufficiently sparse in each image. No other prior information is assumed. Explicit measures are taken to reduce the computational cost of solving the L1 problem. Spatially, a coarse-to-fine strategy based on random subspace projection is employed. Temporally, the smooth temporal variation of the camera motion is exploited.
Find out more about our registration methods in our publications page and via our detailed PDF.