Attentive Mobile Robot Visual Maps via Bubble Memory

Attentive robots can perform efficiently by allocating limited resources to interesting parts of a scene. Fixating on interesting points is achieved by controlling the saccadic camera movements. Attentional sequence is a sequence of features extracted from fixated regions during attentive processing. A model --bubble memory-- was previously proposed for encoding this spatio-temporal data based on deformation of a spherical three-dimensional surface. The bubble surface can be deformed according to the feature values observed at the fixation points. This paper develops bubble based visual map-building for an attentive mobile robot as it is navigating. At each viewpoint, the robot saccades around the scene and deforms a set of bubbles based on its response to a variety of edge, color and corner filters. The bubbles altogether constitute the visual spatio-temporal map as seen from that viewpoint. As the robot navigates to another viewpoint, the bubble set is deformed  according to the new sequence formed there.  It doesn't need to store the newly formed bubble set or search in its memory unless major changes due to scene changes have occurred. Experiments reveal that viewpoints having similar visual scenes have bubble sets which are also similar.