Re-mapping of visual saliency in overt attention: A particle filter approach for robotic systems

Abstract

‘Saliency map’, a scalar, two-dimensional representation of visual saliency, is at the center of almost all of the existing models of visual attention. Attention is directed toward the most salient location in the saliency map. The mechanism of ‘inhibition of return’ (IOR) suppresses the saliency of recently attended location and thereby enabling the shifts of attention toward different locations in the saliency map in order of decreasing saliency. This process performs well as longs as the attention is directed covertly. For overt attention with head movements, which is practically the case in robotic applications, the visual saliency as well as the frame of reference in which the IOR is expressed change after every head movement. These pose a set of computational challenges in implementing attention behavior and IOR. This paper argues that the re-mapping of visual saliency and dynamic shift of IOR emerge naturally in a particle filter based framework of visual attention. Experiments on a real camera head validate the arguments.