COMPUTER VISION ALGORITHMS TO RECOGNIZE AND LOCATE PARTIALLY OCCLUDED OBJECTS (POLYGON APPROXIMATION, CURVATURE ESTIMATION, ASSOCIATION GRAPH, POLYGON FRAGMENT MATCHING)
Abstract
Computer vision algorithms are presented that are used in the recognition and location of partially occluded objects. The scene may contain unknown objects that may touch or overlap giving rise to partial occlusion. It is assumed that the objects are rigid and planar. The algorithms revolve around a generate-test paradigm. Hypotheses of objects are continually generated and tested for visibility in the scene until all the objects in the scene are identified. To generate hypotheses, the boundary of the objects and of the scene are represented by polygons. The polygon representation turns out to be a powerful representation in all phases of hypothesis generation and verification. Special vertices of the polygon are identified as "corners." These corners are used to detect and locate the model in the scene. Polygon moments are used to find the similarity between scene and model corners, and to find the location of the model corner in the scene. Using these matches, a set of mutually compatible matches is used to generate a hypothesis of a model in the scene. The matches and compatibility constraints are represented as an association graph. From the association graph the largest set of mutually compatible matches can be found and used as a hypothesis. The hypothesis gives the location of the proposed model by a coordinate transform that maps the model onto the scene. Hypothesis verification is done by checking for region consistency. The union of two polygons and other polygon operations are used to measure the inconsistency of the hypothesis. The relationships between the objects, such as which object is on the top of the other, is hypothesized by finding the intersections of the verified model polygons. An edge detector, applied to the gray level image representing the scene, is used to determine the presence of the intersecting line segments. Experimental results of all phases of the recognition and location of the objects are also presented.
Degree
Ph.D.
Subject Area
Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.