On the automated construction of image-based maps

For many tasks, we wish to record or recover the description of a remote en vironment so that it can be inspected by a person. This is the problem we a ddress in this paper. Rather than recovering a geometric description of an environment, as many robotics systems attempt to do, we seek to recover a m odel of an environment in terms of its appearance from a set of carefully s elected viewpoints. Our hope is that this type of model is both more access ible to humans for many realistic tasks, and also more readily achieved wit h automated systems. These viewpoints are locations in the environment asso ciated with views containing maximal visual interest. This approach to envi ronment representation is analogous to image compression. Our goal is to ob tain a set of representative views resembling those that would be selected by a human observer given the same task. Our computational procedure is ins pired by models of human visual attention appearing in the literature on hu man psychophysics. We make use of the underlying edge structure of a scene, as it is largely unaffected by variations in illumination. Our implementat ion uses a mobile robot to traverse the environment, and then builds an ima ge-based virtual representation of the environment, only keeping the views whose responses were highest. We demonstrate the effectiveness of our atten tion operator on both single images, and in viewpoint selection within an u nknown environment.