Recognition of partially occluded objects using fuzzy logic

This thesis presents a model-based fuzzy system for the recognition of partially occluded objects. The proposed system consists of two subsystems: a feature extraction and a fuzzy inference system. The first part consists of the thresholding, boundary extraction, gaussian smoothing and the polygonal approximation processes. Local dominant points extracted from the polygonal approximation of the objects contour boundary are used to build characterizing objects features. These features, which are the internal angle and the length ratio, are invariant to rotation scale and translation. An off-line knowledge base of model objects is first built using the described above subsystem. The fuzzy inference system evaluates the compatibility between the candidate and the model objects based on a matching criteria. This matching criteria evaluates the closeness between the candidate and the model objects features using fuzzy terms such as: very small, small, medium and large. The similarity between model and candidate object is expressed using fuzzy terms such as: not similar, not quiet similar and not similar. The system is tested for the recognition of objects comprising composite scenes of industrial parts such as pliers, wrenches and promising results were obtained. The performance is expressed by a controllable relative matching rate that compares the number of 3 or more consecutive matching points to the number of points in the model objects