Studies On Video Saliency Detection Based On Depth Information

Human beings can easily perceive the spatial information of the world around us,at the same time,letting the computer have a similar visual attention as humans is the dream of humans themselves.Simulating the behavior of human visual system,visual saliency detection can automatically generate saliency maps of the target images or the video sequences,by which the computer can be more efficient and more intelligent in image processing and video processing.Visual saliency detection is widely used in many fields,such as image or video representation,object detection and classification,behavior analysis,robot control and so on.Saliency detection mainly consider interesting regions in images and prominent motions in video sequences.Accordingly,the existing two-dimensional based saliency detection methods have achieved abundant research results,but most of these methods compute saliency based on feature contrast with respect to color,shape,orientation,texture,curvedness,etc.When salient objects do not exhibit visual uniqueness with respect to these visual attributes mentioned above,it becomes challenging for existing methods to detect.Stereopsis plays a very important role in human visual system.At the same time,depth information plays a very important role in saliency analysis and the prediction of human eye.In this paper,we study on the depth-based saliency detection for binocular image and binocular video.We provide a new depth saliency method based on the selective difference by analyzing the depth characteristics of binocular images.Then,an improved salient region detection method basd on motion contrast is provided,which is used to detect the temporal saliency of video sequences,at the same time,the depth-based saliency method mentioned above is used to detect the spatial saliency of video sequences,the results of temporal and spatial saliency methods are fused together to obtain the final saliency map.Finally,the saliency detection system for binocular video is presented.The main work of this paper is as follows:Firstly,we provide a new depth saliency method based on selective difference.According to the depth-based saliency detection method,fully excavate the characteristic of the depth information,we calculate the selective difference for global range and local range,and the different values are fused to obtained the final saliency map.Finally,the effectiveness of the method is tested by experiments.Secondly,we provide an improved salient region detection method based on motion contrast.After a detailed analysis for the salient region detection method based on motion contrast,we provide an improved salient region detection method based onmotion contrast to solve the problem that the target area are not detected effectively.Then,we provide a salienct object detection model for binocular video.The improved salienct region detection method are used to detect the temporal saliency for video sequences;the provided depth saliecy detection method are used to detect the spatial saliency for video sequences;the saliency maps obtained by temporal and spatial detection method are fused together as the final saliency map of the video frame.The effectiveness of the model is finally verified by experiments.Finally,we design and implement a saliency detection system for binocular video based on the basis of the above work.We introduce the system design and present the system function at first,then the experimental results of the system are evaluated.