Yet another fast and robust feature-based motion estimation and background layer extraction

In this paper, we describe a method for estimating and stabilizing a dominant motion between two images by 2D pair-wise image registration. Pair-wise image registration estimates 2D transformation parameters between consecutive frames. The transformation between non-consecutive frames is computed by the concatenation of the pair-wise transformations. This pair-wise registration is based on 1) hierarchical parameter estimation and refinement 2) feature-matching 3) FFT-based global matching 4) RANSAC-based parameter estimation. After motion estimation and stabilization, we extract static/non-static components of the video stream and represent them as background/foreground layers. The layer extraction is based on color distribution and netflow analysis. The extracted layers provide a basis for a compact description of a video. The presented approach is illustrated by a set of challenging examples.