Research On Foreground Object Extraction Based On Graph And Depth Layers

With the rapid development of computer vision technology,more and more applications like augmented reality,virtual reality,intelligent monitoring,feature recognition,etc.are used in real life,they bring us great fun and convenience.As foreground objects extraction is the basis of many computer vision applications,digital image processing researchers paid more and more attention to it.Despite the excellent methods proposed recent years,there are still problems such as extractions with unclear silhouettes,incomplete individuals,etc.,practical application is still some distance away.With the appearance of the low-cost depth sensors like Kinect devices,combining color and depth information to extract foreground objects is a novel approach.For the above problems,this paper proposed a method to extract foreground objects in indoor scenes based on graph and depth layers.First,this paper reviewed the research background,significance and status of foreground extracting,described the existing main problems,and introduced the content of the paper.Then,we explained the key algorithms contained in our proposed foreground object extracting method.Finally,we analyzed the experimental results of our approach,and compared with several other state-of-the-art methods to prove the performance.In this paper,we combined the color and depth information to improve the graph-based method to provide a more over-segmented result of the color scene for region merging step.Then we applied a depth map restoration algorithm to fix the holes and eliminate noise.Next,we used multi-threshold Otsu method to segment the depth map to layers to break the continuity between the foreground and background and the foreground objects themselves based on the assumption that indoor foreground objects vary within a small range in depth.An automatic depth layer selection scheme is designed to reduce user interactions.In order to solve the problem that it is unable to find a border where the color and depth is both similar between the foreground and background based on the mentioned steps above and improve the extractions,the depth map is transformed to a surface normal map to give a new constraint.At last,we merged the over-segmented result based on the constraints which are constructed by the depth layer,surface normal map,user specified seed points and area threshold to extract the foreground objects with clear silhouettes in both color and depth.We used the results generated by our approach to compare with the results of several other state-of-the-art methods,experiments demonstrated that our approach is more robust,requires less interaction,the extracted objects are more complete and the silhouettes is clearer.