The Study Of Video Saliency Detection

Human visual attention mechanism can focus on significant stimulus with limited cognitive resources and ignore unimportant information.Visual saliency detection,which simulates visual attention mechanism,aims to find the most attractive regions in the scene.Video saliency detection is widely used in target recognition,target tracking,human-computer interaction and other fields.So,video saliency detection attracts more researchers interesting,and it has research significance and practice-value.This Thesis is concluded as follows:The traditional video saliency detection algorithm is sensitive to the light and complexity of background.To address this problem,we propose a video saliency detection model based on eye-movement information matching and LDP optical flow.The eye-movement information matching LDP features are introduced to the traditional variation optical flow algorithm,which effectively reduces the impact of light on dynamic characteristics.Then,the dynamic characteristics obtained from the improved optical flow are fused with the color,histogram,texture and other static features at the super pixel level.Finally the corresponding SVR model is used to get saliency value for each test sequence superpixel.The experimental results show that this method has better performance on video saliency detection compared with the several state-of-arts.For the video scene detection with the small objects which move fast,the previous method can not extract enough positive samples to train the SVR models.So,we propose a spatiotemporal saliency detection method based on Bayesian model.In this method,the static and dynamic saliency maps are fused to obtain the video saliency map by the improved Bayesian model.Experimental results demonstrate that our method can detect the small objects with fast moving accurately.For saliency detection in complex scenes video,the eyemovement information dataset for complex scene is constructed.The statistic features available and dynamic feature are extracted according to the fixations of hunman being.The samples are trained by Adaboost method and the final saliency region can be obtained.Experimental results illustrate that our method can detect saliency region effectively in complex scenes.