Estimating dense motion fields using a local two-component model

This paper describes a region based approach to computing dense estimates of 2-D motion fields in image sequences. By dense we mean that we would like to have an estimate of the motion at each pixel location at any point in the sequence. In comparison with alternatives, such as optical flow methods, region based approaches to this problem have a number of advantages. Since they involve correlating regions of pixels, they tend to be more robust in the presence of noise and, if the region window is chosen carefully, they can overcome aperture problems. However, previously these advantages have been limited by the simplicity of the underlying model assumed, i.e. typically a single translational motion. Since in general the latter is only valid within small regions, the gains in robustness can be significantly reduced. We adopt a two-component model of local motion-the change between corresponding regions in adjacent frames is assumed to be caused by the presence of either a single motion or two separate motions. The latter case would typically correspond to a motion boundary caused by a difference in depth, i.e. the region can be considered to consist of two sub-regions corresponding to each motion. In addition, we assume that the global motion field can be described by a set of such regions selected from a quadtree tessellation, i.e. blocks at different spatial resolutions.