A FEATURE CONSTRUCTION METHODOLOGY FOR RECOGNITION OF COMPLEX PATTERNS IN SCENES

A framework and techniques for recognition and analysis of complex patterns in scenes are presented. By a "complex pattern" we mean a configuration of objects of interest, called "simple patterns", which appear in full or partial view in a scene. The recognition and analysis of a complex pattern are achieved by identifying all its simple patterns, and then ascertaining how these simple patterns are configured into a complex pattern by determining their relative position and orientations. Some of the difficulties that complicate the problem are the confusion caused by objects of no interest ("clutter"), by occultation of some objects by others, and by deformations caused by the viewing position and noise. We identify simple patterns by comparing them to the scene in an appropriately constructed feature space. Our approach permits us to cope with the above difficulties. Our main contribution lies in the novel way we construct the above features. The structure of each simple pattern consists of a number of "local configurations". The features are derived from these local configurations. They (the features) are "invariant" with respect to a group of transformations determining relative positions and orientations of simple patterns. They are also "complete" (in an appropriate sense) and "stable" (with respect to local deformations). The ensuing mathematical developments lead to a recognition and analysis algorithm which has been successfully tested by computer simulations