Nonrigid Object Recognition against Complex Background Based on Distributions of Local Patterns

Abstract

Stable feature extraction is needed to recognize nonrigid objects. Color Constant Color Indexing(CCCI) is a feature robust against deformation and illumination change, but its stability fails if the domain of the object is not segmented suitably. This is because the domain of the background is blended. In general, object segmentation is difficult if the background is complex. The method proposed herein uses a collection of local images and recognizes the object based on the distribution of the collection in the CCCI feature space. High performance recognition is achieved without object segmentation because we use the statistical characteristic that the feature vectors extracted from the background domain lie near the center of the distribution in the subspace of the object.