Image sequence segmentation using mixture models with uniqueness and spatio-temporal consistency constraints

Operations such as video coding and digital video restoration rely on knowledge of the motion that has occurred in the scene. Whereas research in the 1980's and early 1990's concentrated on reliable recovery of motion fields on a per-pixel or per-block basis, the current trend is towards object-based motion estimation, and the aim is simultaneous recovery of the significant motions in a scene and the corresponding regions of support. Mixture-model techniques are ideally suited to this task. They are described in the following sections, some extensions are proposed, and consideration is given to issues such as modelling of occlusion and uncovering, determining the relative depth of the objects in a scene, and estimating the number of objects in a scene.