A Novel Visual Saliency Model for Surveillance Video Compression

Human visual system is very fast at detectingsalient information of a scene. This detection mechanism ishardwired into our HVS. In many applications there is aneed to find a robust visual saliency detection method thatmimics this detection mechanism in the human visual system.Several saliency models are proposed in the literature mostof them ignore the dynamic aspect of the saliency. Thus weneed a dynamic saliency model that accounts for the static aswell as dynamic information of the scene in order to computethe overall saliency. In this paper we propose to combine apredictive visual saliency model with static features, motionfeatures and face detection into a single model. We introducea new approach to compute saliency maps for videos usingback-propagation neural network, salient motion informationand prediction. The proposed model is tested and validated forsurveillance videos.The similarity of the saliency maps with experimentallyobtained gaze maps is evaluated both visually and with quantitativemeasures. We also propose a compression model forH.264/AVC videos based on this visual saliency model. Ourexperimental results show that we can encode videos withhigher quality than those obtained with the standard baselineprofile of the JM 18.0 reference encoder, while producing onlyslightly larger files.