Representing and identifying jointed objects using a multiresolution technique

Abstract

A multiresolution technique for representing and identifying objects in images is presented. Images are pre processed to extract object contours. Extracted contours are parameterised to form 1-D arrays. Object representations are then constructed as sets of contour waveforms viewed at different resolution levels. In the learning phase of the proposed algorithm, a database of object models is constructed which representations of objects obtained form images. The database includes jointed objects defined as those consisting of more than one part. During the object identification phase, object parts are allowed to be moved in the plane perpendicular to the joint's axil. To make the algorithm fast and efficient, only a subset of the representation points is used for object identification. The performance of the proposed algorithm is investigated under shifts, rotations, and scale changes.