Recent advances in motion interpretation based on image sequences

Abstract

ION IN MOTION INTERPRETATION Natural language provides the means to describe a complex development at various levels of abstraction. A hierarchy of abstractions from image sequence data towards natural language notions is employed to organize the references. In the following we assume that only gradual changes occur between consecutive frames of a sequence. Otherwise, just the kind of scene—specific knowledge which we desire to extract from the observations would already be required to bridge the gaps in the input data. Low—Level Symbolic Descriptions A first level of abstraction is the localization of greyvalue variations characteristic enough to be found again in a subsequent frame. Examples are zero-crossings — i.e. locations of large greyvalue gradients (6) —, .heu±'istic "points of interest" (7, 8), corner points (9,10,11), edge segsients (12), or primitive regions including those used for cross-