Layered motion estimation

Abstract

A layered motion estimation algorithm is proposed that permits quasisimultaneous motion estimation/segmentation up to a fixed maximum number of layers. The estimation results in one motion parameter set per layer, and a segmentation map that assigns these sets to different parts of the image (motion layers). Motion in a layer is modelled with at maximum four parameters capable of describing pan, tilt and zoom. The concept shows some hierarchy, i.e. a ranking of the motion layers. In this way the motion parameter estimation concerning one layer excludes those parts of the image that have been described by a layer ranked higher in the hierarchy and are not polluted by parts of the image that are better described by layers ranked lower in the hierarchy. The concept results in a very low operations count. It has been shown to perform well even in critical scan rate conversion applications, particularly in picture rate up-conversion. A variant including three layers has been scheduled to run in real-time on a Philips TriMedia processor.