An expected motion information model of salience for active cameras

Abstract

Faced with the challenge of learning about its environment, an important first step for a robot is deciding how to direct its sensors to extract meaningful information. Computational models of visual salience have been developed to predict where humans tend to look in visual scenes, and thus they may provide useful heuristics for orienting robotic cameras. However, as we show here, current models of visual salience exhibit some important problems when applied to active robotic cameras. Here, we describe a new model of visual salience, named EMI, specifically designed to work on robotic cameras. The intuition behind this model is that, regardless of the task at hand, it is critical for robot cameras to keep track of motion, be it caused by camera movement or by world movement. Thus, it is reasonable for cameras to focus on image regions that are expected to provide high information about future motion, and to do so in a way that blurs the image as little as possible. We show that EMI predicts human fixations at least as well as current models of visual salience. In addition, we show that EMI overcomes the limitations that current visual salience models have when applied to robotic, active cameras.