Shape Understanding Via Fuzzy Models

Abstract

This paper deals with part of an interface system between a planar robotics scene and a human operator who provides imprecise verbal descriptions of objects included in the scene. Namely the problem of the shape analysis of a segmented contour, in terms of geometrical categories, as used by individuals, is considered. The shape analysis is performed by decomposition of the original contour into several levels of details corresponding to non-convex parts of the contour. A fuzzy labeling procedure then calculates a degree of matching between a convex outline and ideal standard geometrical shapes. Several types of models are used for the description of objects, according to whether they are convex or not, contain more than one region or not. Shapes are identified without the usual preliminary learning stage, and on the basis of non-contextual, dimensionless features.