Research On GPU-Based Parallel Algorithm And System Of Virtual View Rendering

With the technology progress, the FreeViewpoint Video (FVV) technologyallows users to interactively select any viewpoint angle and control3D stereoscopicvideo that can produce realistic visual world, and then can view3D realistic scenesfrom multi-angles. The FVV system is currently a hot research and is the inevitabletrend of the future development of multimedia technology.Generate a new virtualviewpointin free viewpoint video system can be divided into the MBR (Model BasedRendering)method based on3D reconstruction and IBR (Image-Based Rendering)method. Generally, IBR method is mainly implemented by DIBR (DepthImage-Based Rendering) algorithm, DBR (Disparity-Based Rendering) method andRBR (Ray-space Based Rendering) method.The DIBR technology includes cameracalibration, coordinate mapping,occlusion handling and hole filling. The depth map ofreference image is the main factors to affect the image quality of virtual viewpoint.This thesis studied the depth map preprocessing technology based on imagefilteringand the relationship between depth map preprocessing and destination imagequality.Due to thecase that traditional3D-warping algorithm need to preserve theprovisional image and depth map in calculation procedure, and synthesize the targetimage with two times loop access to the image matrix, and not only take moretime,but also need extra memory space to store median results, so it is definitelyinefficient in rendering period.In order to improve the computing efficiency,thethesisdevelops an improved optimization method based on the depth informationfor the virtual viewpoint rendering. The improved algorithm take the mapping flagmatrix method to solve time-consumingproblem and memory overhead of the largercalculation task from the traditional method. The experiments show that theoptimization algorithm save computing time and does not reduce the PSNR value.With the advantage of saving memory space, the optimization algorithm is moresuitable for the GPU parallel computing technology that can improve the computingperformance of the algorithm,Therefore,the improved method improve computing efficiency by reducing the computing time.In order to pursuitmaximal computingefficiency and faster calculation speed, thisstudy also focus on the GPU-based heterogeneous parallel computing technology, itcontains GPU thread design strategy, data asynchronous transfer between the GPUand CPU, and memory access strategy.Alsothe GPU-based parallel virtual viewrendering optimized algorithm is proposed here,and the experiments show that thenew parallel algorithm has the better parallel speedup.Finally,the research focus onstereo math and disparityestimate for virtualviewpoint rendering technology, pre-processalgorithm for depth map, andstereoscopic display principle. Implement a virtual viewpoint rendering system whichprovidesinteractive operation for user. This system has3D image and video capture,stereo parallax estimate, camera calibration, depth map preprocessing, the imagesequence and video sequence reading, the camera parameter configuration, freevirtual viewpoint generation,3Dstereo display and GPU-accelerated renderingalgorithm. The system integrates the basic algorithms of virtual viewpoint renderingtechnology, and the system can be applied to the teaching and experiment platformoftheFVV system, it has significancemeaning to the further study of the practical FVVsystemand GPU-based parallel optimized accelerated rendering technology.