Predictive visual saliency model for surveillance video

Abstract

Visual saliency models(VSM) mimic the human visual system to distinguish the salient regions from the non-salient ones in an image or video. Most of the visual saliency model in the literature are static hence they can only be used for images. Motion is important information in case of videos that is not present in still images and thus not used in most of VSMs. There are very few saliency models which take into account both static and motion information. And there is no saliency model in the literature which uses static features, motion, prediction and face feature. In this paper we propose a predictive visual saliency model for video that uses static features, motion feature and face detection to predict the evolution in time of the human attention or the saliency. We introduce a new approach to compute saliency map for videos using salient motion information and prediction. The proposed model is tested and validated for surveillance videos.