Partition projection in videos by global and local block-matching

The paper presents an original partition projection method for video sequence segmentation. Let us consider that the partition P(t) of the frame I(t) is already available. The aim of the method is to detect the local changes between frames I(t) and I(t+1). These changes are processed as uncertainty areas when initializing P(t+1). These areas are segmented and refitted to the unchanged parts to give the final partition P(t+1). The detection of uncertainty areas is achieved by two block-matching processes. First a global block-matching is used to detect drastic changes between two successive frames. Then a local block-matching algorithm, applied on specific blocks centered on the partition borders, allows uncertainty pixels in the neighbourhood of the new borders to be declared. The local segmentation of uncertainty areas is facilitated by using irregular pyramid segmentation based on an adjacency graphs representation.