MULTI-SCALE CONTOUR SEGMENTATION AND OCCLUDED OBJECT RECOGNITION

Abstract

Contour segmentation plays an important role in occluded object recognition. In this paper, we propose a segmentation method which is robust with respect to noise. Our approach is based on the curvature zero-crossing points extracted from outermost contours of mulit-scale images. By using the segmentation scheme, a matching procedure for object recognition with occlusion is presented. This requires an affine integral invariant representation of contours which have an arclength as parameter.