Motion identification from image sequences: Information theory and pixel selection

The paper deals with the issue of (possibly nonrigid) motion identification from a sequence of degraded images. In particular the central problem is the one of proper pixel selection: the information content of each pixel is the basis of the pixel selection procedure. After the general setting of the problem we give some identifiability conditions. Then the pixel selection procedure is presented. Finally, the procedure is validated against simulated images of moving objects. It appears, that is enough to process few percent of the total pixel number without any significant decrease of the estimation accuracy.