Optical Flow Using Overlapped Basis Functions for Solving Global Motion Problems

Abstract

Motion problems in which the scene motion largely conforms to a low order global motion model are called global motion problems, examples of which are stabilization, mosaicking and motion superresolution. In this paper, we propose a two-step solution for robustly estimating the global motion parameters that characterize global motion problems. Our primary contribution is an improved estimation algorithm for modeling the optical flow field of a sequence using overlapped basis functions. Moreover, we show that the parametrized flow estimates can be consolidated through an iterative process that estimates global deformation while ensuring robustness to systematic errors such as those caused by moving foreground objects or occlusion. We demonstrate the validity of our model and accuracy of the algorithm on synthetic and real data. Our technique is computationally efficient, and is ideally suited for the application areas discussed here, viz. stabilization, mosaicking and super-resolution.