On-line unsupervised planar shape recognition based on curvature functions

Abstract

In this paper, a correlation based method for planar shape recognition is presented. First, shapes are represented by its contour chain code. Then, two local histograms are calculated on both sides of each point of this code and correlated to obtain a curvature function associated to the shape. This function is normalized by decimation or interpolation to a suitable fixed value and supplied to a classification stage. The classification stage consists of an unsupervised clustering process grouping by similitude, defined as the maximum value of the correlation of the input object with the prototypes of the classes conformed by the previous classified shapes. The clustering process often recalculates the correct number of classes with almost no additional time requirements: thus, it creates new classes to deal with new objects and fuses old ones if it is necessary. The proposed systems works on-line with no previous training process and is spacially suitable to be used by autonomous mobile robots working in real time in unknown environments.