Detection of dominant points on an object boundary: a discontinuity approach

Abstract

In a typical partial shape recognition scheme, each scene or model object is decomposed into a concise representation formed by a sequence of contiguous primitive features terminated by pairs of dominant points. A scene object is classified as a model object if its representations are similar. The main problem of this approach is caused by the inconsistency of the dominant point distributions as images of the same object are taken under different levels of noise interference and spatial variations. In this paper a robust dominant point detection algorithm is presented. The technique employs optimal discontinuity detectors for locating the structural nodes of an object boundary with high noise rejection and accurate localization capabilities. Members of a structural node include corners and the terminals of lines and arc segments which are unaffected by spatial variations. The algorithm has been tested with images of a set of hand tools grabbed under different scaling, camera orientations and noise interference. For the majority of cases, a similar set of dominant points is obtained for different images of the same object. The encouraging results demonstrate the feasibility of the approach and its potential as a reliable basis for object recognition.