Development of a segment-based description of events in image sequences

Abstract

This paper describes an approach for image sequence processing, which extracts objects (or parts of objects) from the image sequence, generates motion-specifically attributed tracks (location/time course) and derives a hierarchy of motion. The base consists of object-characterizing segments which are determined by a multiple threshold method. This method yields feature descriptions for each segment. Based on this description, an association of the segments from frame to frame is done. The resulting motion vectors are concatenated as long as the features of the segments and the motion-specific attributes do not change significantly. In a subsequent step the analysis of motion-specific interrelations between the tracks is executed. The event recognition is processed and controlled by event models. In conformity with motion concepts the hypotheses of events are initialized by motion dependent changes of the tracks.<>