Research Of Video Saliency Detection Algorithm Based On Optical Flow

As the video can carry a wealth of information,the applications of digital videos have been more and more widely in the daily life.When watching a video,large amounts of information will get into the field of vision,but only a few are useful.The human visual system typically captures prominent areas and movements in the video to reduce the search effort.It can be seen that the visual attention mechanism is the inherent property of the human eye.The area where the human eye fixed is the salient region.To predict the salient area in the video is of great significance in many applications,such as monitoring system,target recognition,target tracking,video quality evaluation and so on.Video saliency detection plays an important role in video processing.Based on the requirement of video quality evaluation,a video saliency detection method based on optical flow is proposed in this paper.The new method is based on the existing image detection algorithms.By calculating the optical flow,we obtain the significant information of the video in the time domain.And combined with the salient area of the video in the spatial domain,we can calculate the video saliency.Then we also use the eye fixed results to assess the performance of the algorithm we propose.The main work of this paper includes the following three parts.Firstly,the paper summarizes the current research status of visual saliency detection at home and abroad and some basic characteristics of the visual attention mechanism and the image,which provide the theoretical basis for the research of the saliency detection algorithm.Secondly,aiming at the two aspects of current video saliency and video quality evaluation,this paper designs the eye movement data acquisition experiment in order to provide the ground truth of salient regions in the video where the eyes truly fixed.With the video quality evaluation database,such as LIVE,VQEG and so on,as the video source,the Tobii X120 eye tracker is used to collect the fixation position of the human eye while watching the video.At the same time,based on the collected eye tracking data,the gaze graph and the hotspot graph on the video frame are calculated by statistical analysis to make the eye movement data visualized,and the video saliency database based on the eye tracking data is constructed.Thirdly,because the motion is the expression of the information of the video in the time domain,and the optical flow can estimate the motion vector in the video,this paper introduces the optical flow in the video saliency detection on the basis of the existing image saliency algorithm.The information contained in the video is decomposed into two parts in this paper,that are video frame information on the spatial domain and the optical flow map information in the time domain.In this paper,we extract some features both in video frame and the corresponding optical flow map,and analyze their salient region,respectively.And the final video saliency can be calculated by merging the results of the spatial video saliency and the temporal video saliency according to the global motion analysis.Finally,some videos are selected on the eye tracking database to evaluate the performance of the algorithm we have proposed.The experimental results show that compared with some existing algorithms such as GBVS and so on,the results of our algorithm are in better agreement with the actual results of human eye.