Human visual field based saliency prediction method using Eye Tracker data for video summarization

Abstract

Video summarization is the process to extract the most significant contents of a video and to represent it in a concise form so that a user can understand about all the important events of a long video. Existing methods for video summarization fails to achieve a satisfactory result for a video with camera movement and significant illumination changes. To solve this problem, a new saliency prediction method is proposed based on human visual field using human eye's fixation data obtained by Tobii X120 Eye Tracker. Three different circular regions are considered around a fixation point similar to foveal, parafoveal and peripheral regions of human visual field. The inner (foveal), middle (parafoveal), and outer (peripheral) regions are assigned highest, mid and lowest salient values respectively. The motivation is that human pay more attention in foveal region and less attention in parifoveal region. Based on this concept, a visual saliency score is calculated from eye tracker fixation data for each frame and a set of key-frames are selected based on used preferences. The proposed method is implemented on Office video dataset that contains video with camera movement and illumination change. Experimental results show superior performance compared to the existing GMM based method.