nextuppreviouscontents
Next:View-Based Route-Learning Up:Active Perception Previous:FPGA-Based Spiking Neural Networks

Attention and View-Based Object Recognition

This project involves the application of attentional processes to the implementation of view-based object recognition schemes. In this approach, objects are recognized by comparing their appearance from a given viewpoint to (interpolated) views which have been previously recorded. Attention promises to improve this process by minimizing irrelevant inputs and by increasing viewpoint invariance. In such an active view-based object recognition technique, hypotheses as to an object's identity alter the way in which further information regarding the object is gathered. Previously, we had implemented the low level attentional shift algorithms based on hard-wired measures of saliency (colour and saturation), and have used these to guide the motion of a video camera via a Directed Perception Pan-Tilt unit. In the last year, we have developed an object representation that can be used to integrate and learn the association between observed views and object identities. We are currently testing our algorithms in a demonstration environment.

A. Domercq, J.J. Clark
 


Annual Report

Mon Jun 26 21:22:20 GMT 2000