Proceedings of the Cognitive Robotics Seminar

Visual attention is the biological mechanism allowing to turn mere sensing into conscious perception. In this process, object-based modulation of attention provides a further layer between low-level space/feature-based region selection and full object recognition. In this context, motion is a very powerful feature, naturally attracting our gaze and yielding rapid and effective shape distinction. Moving from a pixel-based account of attention to the definition of proto-objects as perceptual units labelled with a single saliency value, we present a framework for the selection of moving objects within cluttered scenes. Through segmentation of motion energy features, the system extracts coherently moving proto-objects defining them as consistently moving blobs and produces an object saliency map, by evaluating bottom-up distinctiveness of each object candidate with respect to its surroundings, in a center-surround fashion.