| This dissertation deals with the problem of identifying partially occluded objects from a vision image. Three major procedures are proposed: location of cornerpoints to represent images, partial matching of a given set of images, and hypothesis testings.;In locating the cornerpoints of images, an optimization-based derivative-free line search method was developed to locate all the cornerpoints of 2-dimensional polygon image for pattern recognition and inspection purposes. This optimization-based method was also used to approximate a 2-dimensional non-polygon object by a polygon to any desired accuracy. The computation time of the proposed method is less than 10% of that required by other well-known methods.;In the partial matching procedure, a graph-theoretic optimization method is used to recognize partially occluded objects from a 2-D image through the use of maximum curvature points. Maximal cliques and a weight matching algorithm are also developed. The vertices of an occluded object image with high curvature values are classified by the objects which are hypothesized to be involved in the occlusion. A heuristic method is also developed to further improve the computational speed. Typical examples are given to illustrate the accuracy of the optimization model as well as the simplicity of the companion heuristic method.;In the hypothesis testing procedure, a pattern matching scheme is developed to inspect non-occluded objects in an industrial environment. The inspection includes dimensional verification and shape matching which compares a 2-dimensional image of an object to a pattern image. A heuristic axis of orientation which has almost the same rotation accuracy as the axis of least inertia is also introduced, but it has a time requirement which is less than one twenty-fifth of that needed for the axis of least inertia method. The method proves to be computationally efficient and accurate for real time application. The methodology of shape matching may be applied to hypothesis testings in the case of occluded objects. A possible extension of this research is to recognize three-dimensional objects with single perspective views. |
