RECOGNITION OF PARTIALLY OCCLUDED OBJECTS

Abstract

The aim of this work is to demonstrate how to enhance the capability of the computer vision to recognize partially occluded parts.

In the thesis, we show how local features are used for object identification. Local features such as comers, holes, lines, curve segments,...etc. depends only on a subset of an object so it is possible to deal with the problem of recognizing occluded objects in the case where some of the local features are visible.

First, a survey of computer vision process is introduced with an emphasis on the related topics, this is followed by discussion of the problems of occlusion. Next, three techniques have been presented, the first uses the holes in the objects as a local feature for radial template matching. The second, uses the Curvature Guided Polygonal Approximation to represent objects in isolated and occluded scene depends on the existence of critical points of high curvature. Third, contour matching for those objects who do not posse critical points or in cases where such critical points are hidden.

In the simulation stage, algorithms and modifications have been implemented and results are summarized. The system is tested on a real set of handtools and mechanical parts. Different techniques are compared with respect to the ability of detection of occluded objects with boundary separated into isolated clusters, relative sensitivity to noise, the ability of recognize objects of different scales (sizes) and rotation, complexity and

computational load.