Attentive vision, scene representation and bubble space

Visual data based environmental representation is crucial for mobile robot applications requiring recognition. Previous work has shown that bubble memory — which is an egocentric approach based on hypothetically surrounding a spherical surface around the robot, to provide a compact representation of the scene from a single viewpoint. This paper proposes bubble space as an extension of bubble model to time-varying robot viewpoint. In a given scene, at each viewpoint, the robot saccades around and deforms a set of bubbles based on its responses to a set of relatively complex visual filters. As each bubble can be compactly represented using double Fourier series, the associated Fourier descriptors can either be stored in its memory or be used to recognize previously encountered scenes as verified by experimental results with an attentive robot.