An accumulative framework for the alignment of an image sequence

Abstract

We consider the problem of estimating global motion in an image stream containing outlier motion, and propose a novel framework to address the stated problem. Rather than following the conventional approach of analyzing pairs of images, we propose an accumulative framework that utilizes all available information to compute the desired alignment of an incoming image. This approach simultaneously aligns each incoming image to all relevant images, thereby exploiting information contained in the observed sequence that may not exist in an adjacent image. We present an algorithm, within the proposed framework, with the following primary features: First, unlike previous approaches, we increase tolerance to outliers by preserving the integrity of information in every frame and use temporal reasoning to weight the estimation. Second, our method inherently accounts for the concatenation error typical to frame-to-frame alignment techniques, identified in [16, 4]. Third, the algorithm we present does not require non-linear optimization of an inordinate number of parameters-the algorithm iteratively solves linear least-squares equations for a small number of parameters.