Optical flow calculation from feature space analysis through an automatic segmentation process

Abstract

The determination of the shiftings of objects in a sequence of images (i.e., optical flow calculation) requires the search for significant matching points in two successive frames. In this paper, the characteristics points considered are the feature vectors of the “blob” regions into which an original image has been segmented. Each blob region is a set of adjacent pixels with the same properties and characterized by temporal variations. A feature vector has been computed for each blob. The main consideration is that this data set more “preserving” and more typical within frame sequences than the conventional characteristics points (identified only by the spatial coordinates) of some particular pixel. The matching between the regions of successive frames is developed in the feature space. In fact, in addition to the spatial coordinates, in this space it is possible to consider further features which are sufficiently “invariant” thus reducing the probability of wrong matches. The matching algorithm is essentially based on Barnard and Thompson's method, suitably modified to solve uncertain situations. Results have been obtained for a noisy 3-D environment; a comparison between conventional approaches and the present one is made and discussed in the paper.