A Method for Partial Shape Recognition

Abstract

As a function of vision system, it is important to identify each object and decide its pose in a scene. When the scene is the silhouette of planar objects overlapping each other, and when the shapes of objects to be identified are given as models, the sequence of these two tasks is called 2-dimensional partial shape recognition. In this paper, we propose a novel method for partial shape recognition. The proposed method uses information of curvature and tangential line direction at each points on digital contours of model objects and a scene. The pose decision algorithm is based on the concept of generalized Hough transformation. To reduce the execution time, contour points data lists are sorted by their curvature. To increase the reliability of results, contour points of models are re-sampled densely in high curvature segments and coarsely in straight line segments. Moreover, to reduce the memory space for generalized Hough transformation, coarse-to-fine analysis is applied. Experimental results show that the proposed algorithm provides valid results of partial shape recognition and takes less than 20 seconds per one model object for recognition on a typical workstation.