The Techniques Of Depth Extraction For Convering2D To3D Video

With the improvement of people’s life quality and the emergence of3D TV and3D Movie, people have more requirement for the clarity of images and video, and3D video becomes more and more popular, so the conversion of2D to3D video has got the universal attention of people. Nowadays there are two mainly key techniques for the conversion of2D to3D video. One is to recover the depth information from2D images, and the other one is to convert the depth map to3D video which can be displayed on the3D display device. This paper focuses on the research of depth estimation from2D video.The paper introduces the research background of3D video and the key techniques and research status of the conversion of2D to3D video. By the introduction and comparison of several common depth cues used to extract depth information, we learn that one depth clue can only be applied to a particular scene, there is not a universal algorithm of extracting depth information, and the depth maps extracted by these algorithms are not very satisfying. According to the present situation of depth estimation as well as the character of most of the existing video sequences often contain relatively static background and serveral dynamic objects, this paper proposes the method which respectively extracts depth information of the dynamic objects and the static background and then fuse them together. The method contains four steps. Firstly, separating the dynamic objects and the static background from2D images. The separating method uses the mixed Gaussian model algorithm to extract the static background, and then extracts the complete dynamic objects with the adaptation threshold background difference method and the biggest connection areas extraction algorithm. After that, in order to get the dynamic objects image with accurate edge, the paper proposes a method combining the improved adaptation mean shift image segmentation algorithm and projection algorithm in space partition. The results in the experiment show that the proposed method can separate the static background and the dynamic objects very well. Secondly, as the depth information can be recovered by the depth cue of motion information, and by the comparison of several motion estimation algorithms on block matching, the paper presents the adaptive diamond search algorithm which has less computation and higher matching accuracy to recover depth imformation of the dynamic objects. However, there are maybe some burrs and block effect in the motion-based depth map, the paper proposes a method based on image segmentation to smooth the depth map in order to eliminate the shortcoming. The smoothed depth map is usually not consistent with human vision, so the linear perspective depth map is proposed to optimize the smoothed depth map. Thirdly, the method fusing the luminance information and the baseline information which are contained in the background image is presented to obtain the final background depth map, and the experimental result shows that the final background depth map is more precise and it has better visuality. Finally, because of the separation of the background and the foreground, this paper adopts the method which fuses the motion-based foreground depth map and the geometry-based background depth map into the final depth map. The experimental result shows the depth map extraction method proposed in this paper is really practicable, and this method can obtain the satisfying depth map.