OOP: Object-Oriented-Priority for Motion Saliency Maps
Anna Belardinelli, Werner X. Schneider, Jochen J. Steil · Publikationen an der Universität Bielefeld (Universität Bielefeld) · 2010
Abstract The ability to attend to motion is paramount for living beings. The human visual system is able to detect coherent motion and select within multiple moving objects the most conspicuous or most relevant to the task at hand. Similarly, any artificial agent operating in dynamic environments needs to be endowed with a mechanism for rapid detection and prioritization of moving stimuli in its field of view. In this paper, we present a biologically and psychologically inspired model of this ability and tune it for the extraction of motion at different scales and velocities. Unlike many computational models that compute saliency pixelwise, we extract moving proto-objects through segmentation of motion energy features. These perceptual units, so called proto-objects, are identified as consistently moving blobs. A proto-object based priority map is hence obtained by assigning a single saliency value to the region confining a segmented object. Priority stems from a combination of bottom-up saliency, evaluated in a center-surround fashion, and from top-down biasing of motion features or motion saliency. Experimental simulations on synthetic displays and real sequences show the effectiveness of the proposed approach.