Research On Video Saliency Detection Method Based On Spatiotemporal Gradient

Video saliency detection is detecting the most attentive region in the video frames.Visual saliency detection is an important step in many computer vision applications,such as image retrieval,image/video compression,video surveillance,since it reduces further processing steps to regions of interest.Saliency detection in still images is a well-studied topic.However,videos scenes contain more information than static images,and this additional temporal information is an important aspect of human perception.Therefore,it is necessary to include motion information in order to obtain spatiotemporal saliency map for a dynamic scene.Dealing with videos is more complicated than static images and the perception of video is also different from static images due to additional temporal information.A video shows a strong spatiotemporal correlation between the continuous frames.A common approach to deal with video sequences is to compute a static saliency map for each frame and then combine it with a dynamic saliency map to get the final spatiotemporal saliency map.The accuracy of the saliency model depends on the quality of both the static and dynamic saliency maps and also on the fusion methods.These methods compute the static and dynamic features of the video frames separately,without taking into account that the video is a whole structure.Generally speaking,moving objects attract more attention,so motion information plays an important role in videos.Static and motion information all contribute to the salient objects.Based on this consideration,a new method combines static and dynamic features is proposed to detect video saliency based on spatiotemporal gradient.The method includes the following steps:First of all,the input video frame is converted into the CIELab and CIEHSV color space to get color features,and the motion information is calculated by an optical flow method.The FCM clustering algorithm is used to cluster in both color and motion channel,respectively.Next,the color gradient and the optical flow gradient of the video image frame are calculated,and the spatiotemporal gradient map is obtained by combining the color gradient and the optical flow gradient.The spatiotemporal gradient value of each clustering region is calculated by the clustering region on each feature channel and spatiotemporal gradient map,and then an adaptive threshold selection method is proposed to identify the candidate salient regions and background regions.And the region with the largest spatiotemporal gradient value in the candidate salient regions is selected as the focus of attention.And then,a fusion process is proposed to get the complete salient regions from the candidate salient regions.Finally,the saliency map is measured by using the color difference feature between the salient region and the background region.A Gaussian filter is used to further optimize and obtain the final saliency map.Experimental results demonstrate that the proposal algorithm detects saliency precisely and reliably,and show that our method outperforms state-of-art methods.