Real-time statistical saliency using high throughput circuit design and its applications in psychophysical study

Abstract

Using low level video data, features can be extracted from images to predict search time and statistical saliency in a way that models the human visual system. The statistical saliency model helps explain how visual search and attention systems direct eye movement when presented with an image. The statistical saliency of a target object is defined as distance in feature space of the target to its distractors. This thesis presents a real-time, full through-put, parallel processing implementation design for the statistical saliency model, utilizing the stability and parallelization of programmable circuits. Discussed are experiments in which real-time saliency analysis suggests the addition of temporal features. The goal of this research is to achieve accurate saliency predictions at real-time speed and provide a framework for temporal and motion saliency. Applications for real-time statistical saliency include live analysis in saliency research, guided visual processing tasks, and automated safety mechanisms for use in automobiles.