Progressive saliency-oriented object localization based on interlaced random color distance maps

Abstract

The human visual system employs an information selection mechanism, visual attention, so that higher-level cognitive processes can be restricted to a potentially important subset of the incoming information. This mechanism is amenable to efficient computational implementation and, consequently, it has been incorporated into many technological applications. Among these applications is autonomous mobile robotics, in which efficient vision systems are paramount given their limited computational resources and energy autonomy requirements. Robots have employed visual attention for directing gaze and accelerating object detection, among other tasks. However, it has always been approached as a single rigid input-output stage. In this work, a bottom-up, unsupervised visual attention model based on progressive processing is presented. It adopts an incremental approach, in which a rough output is rapidly computed, and then successively refined, providing graceful-degradation, which might be particularly useful by robots based on the subsumption architecture. Progressive processing is achieved at a pixel saliency estimation level by adopting a method based on color distance to random pixel samples, and at a scale level through bidimensional interlacing. The proposed approach is assessed in the SIVAL dataset, and compared to other two visual attention models commonly employed in robot vision, presenting highly competitive performance.