Depth Map Acquisition Based On Fast Image Matching And Virtual Viewpoint Rendering

With the development and maturation of three-dimensional display technology, 3D TV and movies gradually come into people’s lives, and the demand of associated stereoscopic videos has greatly increased. Stereoscopic videos can be captured directly, or generated through some post-processing virtual viewpoint techniques. In 3D display, viewing experience could be more realistic by using more viewpoints, but it cost a lot through directly shooting. To increase shooting viewpoint,it will come with much higher cost and greater shooting difficulty. Therefore, common 3D filmmaking normally shots no more than two viewpoints, In this paper, a method of viewpoint virtualization is presented to generate a new viewpoint from two viewpoints by highlighting the stereo matching methods which is used to obtain depth map, virtual view rendering and image restoration based on the depth chart.Firstly, the basic theory of stereoscopic display is researched, including several camera related coordinate representation methods and model. We also studied the stereo matching constraints and key technologies, and introduces several stereo matching algorithms.Then, an improved image matching method based on graph cut is presented to extract depth information from the image pair. Traditional graph cut based stereo matching method have better matching results for large low texture areas and image blocking areas. However, the use of fixed coefficients in the energy function does not suit to video image sequence processing, and the matching process requires to compute matching cost for each node resulted in a slow computing speed.In this paper, by establishing a disparity gradient volume combined with stereo matching constraints, the disparity search range is narrowed and matching speed is faster. The experiment result shows that, the matching speed is increased by more than threefold compared with the traditional method. This paper proposes an energy function with adaptive coefficients, combined with stereo matching constraints to narrow match searching area, to accelerate the matching speed.Finally, a virtual view point rendering method based on the depth chart is presented. in which the virtual viewpoint rendering method DIBR is used. DIBR draws out the new viewpoint of regional deformation and occlusion appears empty,and you need to use the appropriate method to fill in the image stretched. In this paper, the background image is reconstructed by making full use of video sequences information combined with depth map, and image is divided into foreground and background areas to be repaired by different methods so the filling effect can be more reliable. From the experimental results, PSNR data obtained by this method is superior to other algorithms, and compare with the real structural similarity viewpoint reached 99.1%, up more than 10% compared to traditional methods.Experimental result shows that, compared with the traditional methods of virtual viewpoint, the method in this paper by optimizing the matching method, greatly improved the speed of depth chart extraction, and can restore the authentic background image to draw a new point of view which is highly similar to the real point of view.