Research On Depth Extraction For2D To3D Video Conversion

Since three-dimensional (3D) video has more impressive stereo vision effectcompared with2D video,3D video technology has won the attention from all sectorsof the society and has rapidly developed.3D video technologies including what from3D content acquisition to3D display cover many technical methods. Among them,3D content acquisition is the key part of3D video service. Although3D content canbe produced by multi-view shooting and some special equipment, there are still lotsof key technologies under solved, that results to the lack of3D video content andlimits the popularization and application for3D video.2D to3D video conversion isan efficient solution to meet this lack. As the accuracy of depth map may seriouslyinfluence the quality of the reconstructed3D scene, depth map generation is a keystep in the2D to3D video conversion.This paper researches on the depth extraction for2D to3D video conversion.Different depth cues are used to obtain depth information. Detailed introduction isgiven for2D to3D video conversion principle, depth cues based on human visionand depth extraction methods using different depth cues, which are theory basis forour research work. The main research work and innovative results are as follows:Using the role of region segmentation and merging in vision perception, wepresent a depth extraction method based region merging. Firstly, the statisticsprinciples are used to merge regions. Then, according to different scene depthassignment models, we generate prior-hypothesis depth map. At last, combine theimage regions with prior-hypothesis depth map, assign depth values to the imageregions, and generate the final image depth map. For the first time, this methodintroduces region merging into the depth extraction, which can adjust the mergingalgorithm according to the complexity of image scene. In addition, we classify thedepth assignment models into two kinds: plane model and circle model, and expandtypes of the prior-hypothesis depth maps for common scene to seven, which effectively enlarge the application of this method. The experimental results show thatthis method can more accurately extract the depth information from2D image scene.Using multi-cue to2D video depth extraction, we propose a depth extractionmethod based on linear perspective information and motion information. Thismethod mainly consists of two parts: depth extraction based on linear perspectiveand depth extraction based on motion. The depth extraction based on linearperspective uses the image edge detection to detect the vanishing point andvanishing line. This method also proposes a depth assignment strategy according toscene type to obtain the depth information for the static background. The depthextraction based on motion uses motion estimation and motion detection to extractthe depth information for the moving objects. Finally, we propose two differentdepth fusion methods to get the final depth map for the video sequence. Theexperimental results show that the proposed method is effective to depth extractionfor videos sequence with depth cues of linear perspective and motion. The defect ofinsufficient depth extraction from single depth cue is improved by using theproposed method.