Component based recognition of objects in an office environment

Abstract

The Problem: The goal is to develop a trainable system for object recognition which is able to identify common office objects under different pose and illumination conditions in still gray images. Motivation: The need in robotics to recognize common objects for navigation and/or interaction is of increasing importance. In current systems imaging parameters such as pose and illumination still cause much difficulty in correct recognition. Previous Work: Previously, component-based approaches have shown promising results in various object detection and recognition tasks such as face detection [2], person detection [4], and face recognition [1]. Approach: We adopt the component-based classification architecture suggested in [3] for object detection. It consists of two levels of classifiers: On the first level, component classifiers independently detect components of the object. In contrast to [3] where SVMs were used on both levels, we use component templates and normalized correlation for detecting the components and a linear classifier to combine the results of the normalized correlation. In a first step we extract a large number of components from the object images and cluster them based on the similarity of their image features and their locations within the image. The resulting cluster centers build an initial set of component templates from which we select