Research On Reconstruction And Display Technologies Of Three-Dimensional Video Based On Auto-Stereoscopic Display

Three-dimensinal display technology can provide people with immersive3D feelings, so it has always been the research hotspot of the display field. By providing multiview imges of the same scene simultaneously, auto-stereoscopic display technology can support multiple viewers, who can feel strong and amazing3D effects with naked eyes. As the coming of3D times, auto-stereoscopic display technology has been becoming the focus of the institutes and corporations, and is sure to be the next direction of display technology.Based on the auto-stereoscopic display system platform, this paper make research on the fields of virtual view redenring algorithm, parameter design of virtual camera array, texture restoration of disoccluded areas, and multiview auto-stereoscopic display technology. The achievements are listed as follows:(1) A new backward depth-image based rendering algorithm is proposed. The transition areas between foreground objects and background scene in the depth map are detected and marked. Because of their unreliable depth values, the transition areas will be skipped during the rendering procedure of virtual views, which will eliminate the false contours around the foreground objects in virtual view images. Depth map with simple texture will be projected to the virtual view position to generate the virtual depth map, then the virtual depth map is re-projected to the reference view to get the corresponding texture imge of the virtual view. For each pixel position in virtual depth map, there exists a depth buffer, which will keep the depth of foreground pixels and delete the depth of background pixels. The proposed backward depth-image based rendering algorithm can handle the visibility and resampling problems properly, and achieve high image quality.(2) A new parameter designing method of virtual camera array is proposed, which is suitable for the flat-pannel auto-stereoscopic display technology. Theoredical analysis proves that Shift-Sensor camera model is the optimal stereo camera model with best stereo effect and without any geometrical distortions. Then the relation of the parameters is derived for the optimal stereo camera model, and the designing steps of internal and external matrices are demonstrated. The mapping relation between real3D space and virtual3D space with least geometrical distortion is presented. The presented parameter designing method of virtual camera array has rigorous theoretical basis, good interactivity, and ture3D feelings.(3) A new texture restoration method of disoccluded areas combining with temporal and spatial disocclusion information is proposed. Stationary scene information of the input sequence is obtianed by stationary scene extraction algorithm. Then the extracted stationary scene is rendered to the virtual view position to recover the missing texture of disoccluded areas. By doing this, the area of the hole regions will decrease. The texture of the remaining blank regions are restored by the proposed oriented examplar-based inpainting algorithm, which will fill the disoccluded areas in the background to foreground order, and select the best matching block with the most similar texture and depth structure to fill the target block. By effectively using the temporal and spatial disocclusion information, the proposed disocclusion filling approach can obtain convincing and plausible results with improved temporary texture consistency.(4) A new subpixel mapping algorithm with view interpolation for the most popular lenticular auto-stereoscopic display technology is proposed. Firstly, the view number of every subpixel is calculated according to the generalized subpixel mapping algorithm. Secondly, the input multiview images are down-sampled by linear interpolation algorithm to simulate the low-pass filtering process and up-sampled to the full resolution of the auto-stereoscopic display. Thirdly, the depth value of subpixel is calculated according to its assigned view number N, then the depth value is mapped to the disparity value between two adjacent views. Finally, the corresponding pixels in the two nearest integer views of view N are found out. The same subpixel component values of these two corresponding pixels are used to calculate the value of the current subpixel component. The final image multiplexed by the proposed algorithm has fine texture and smooth transition between different view zones.