The human visual system interprets the succession of frames as a moving scene. In this process a pattern recognition stage is involved because the same pattern appears at different locations. Humans get motion attention using our visual “motion-tracker” system. It is natural to think that, in terms of HVS properties, motion detection is more reliable than color vision, shape recognition, or other visual stimulus. One of the main characteristics of any video sequences is the presence of diverse types of motions. Such motions can be produced, for instance, by static objects in a moving background due to camera motion, objects moving in different trajectories, or simply random motion objects. In these scenarios, the human eye gives priority to track fast and unpredictable motion objects (noticeable motion) against slow or predictable motions (non-noticeable motion). Besides, in case of multiple motions in presence, noticeable motion attracts more attention than the rest. It is assumed that for background or non-noticeable motion objects the viewer perceives just the semantic meaning of the displayed objects holding his/her attention to noticeable objects only. Therefore, based on the fact that the HVS gives priority to noticeable motion objects, we want to examine what would happen, in terms of coding efficiency and visual quality, if we developed a video coding approach inspired by the texturebased video coding approach but taking into consideration the assumptions of the human eye motion detectors, instead of texture and spatial properties. To do so we opt for a foreground-background distinction.
Generally speaking, the foreground-background extraction problem aims at estimating the camera movement (global motion) that affects both the moving and stationary points of any video sequence. Extracting the background from a video sequence is still an open problem because a practical algorithm needs to take into account various phenomenons such as illumination changes, background object displacements or non static backgrounds among other effects. Moreover, the method has to be computationally efficient yet robust enough. Various effects may degrade the algorithm performance including moving object shadows that appear on background object displacements, non-static background such as computer screen flickers and so on. A common approach is to model this global motion by a parametric 2-D model. Then, depending on the specific application, the estimated parameters can be used either to compensate the global motion or to extract all those macroblocks, (
16 × 16 pixels), belonging exclusively to global motion patterns. This estimation technique is usually a two-step process: selection of the parametric model and optimizing the process.
