MOTION ESTIMATION BY MEANS OF PRINCIPAL REGRESSION

Abstract. This paper presents an approach where the optical flow between frames of video sequences is estimated according to a pel-recursive strategy through the use of a principal component regression (PCR). This is a simple choice when it comes to treat mixtures of motion vectors due to the fact that it is not essential to have too much knowledge on their statistical properties (although they are supposed to be normal). Local image properties are taken into consideration in order to estimate the 2D motion. The main advantages of the developed procedure are: ( i) the noise distribution is not used directly; and (ii) the types of motion present in neighborhood can be identified, so that a discriminant-type of test can be performed. Preliminary experiments indicate that this approach provides robust estimates of the optical flow.