MODEL-BASED OBJECT RECOGNITION

Abstract

Recognition systems that employ a model-based approach are generally less susceptible to poor illumination and may even be able to deal with some degree of occlusion. This thesis describes the recognition of objects using a model-based method. Two-dimensional objects are represented by boundary segments encoded by sets of codes. Several simple descriptors of the boundary, such as the centroid, extremal boundary points and vertices which are used in the recognition process, are extracted from these codes. These features are used to recognize objects in a model-based scheme. Model database which contains boundary codes and scalar features derived from each of the model objects is constructed. Recognition of an unknown object boundary requires the generation of several hypotheses for testing to determine which is the best possible match. Matching is carried out by aligning the model and object boundaries and determine the match score. The match score that is computed is the percentage of model boundary sections that intersect the object boundary. The recognition system is successfully tested in recognition of objects appearing in unoccluded, occluded, reduced and enlarged scenes. Examples are presented to illustrate each of the scenes.