Abstract

We present a testbed implementation for a foveated robot vision system. To achieve foveation, the visual sensor simulates nonuniform sampling and is mechanically directed toward a specific fixation point. An interest operator is used to select a sequence of fixation points. Successive snapshots of high foveal and low peripheral resolution are combined to create a wide-angle representation of the scene. Such visual processing has been studied with primate and human subjects from a biological point of view. The purpose of this work is to create an active vision system with which we can study foveated machine vision. The current implementation incorporates a CID camera positioned by a PUMA robot to pan and tilt around a fixed point, a SIMD parallel computer, and conventional computers to construct gray-level, edge, and interest maps from several fixations. The system is highly modular, and its architecture permits the efficient incorporation of sequential and parallel components for real-time operation. We demonstrate the modularity of the system and its potential as a testbed for active vision by incorporating two different attentional mechanisms and quantitatively evaluating their performance on artificial and natural images. We propose three types of norms that can be used for this performance evaluation.