Research On Depth Map Based Virtual View Generation

With the progress of science and technology, especially the rapid development of computer vision technology, the public has been able to easily obtain far better visual experience than the past. However, the pursuit of more excellent visual experience from the public has not been slowed down.The new revolutionary vision technology is urgent need to be achieved and popularized. In recent years, many virtual view generation technology based applications have appeared. Virtual view is the non-existent view which is generated through certain algorithms by using of the known view. The virtual view based applications often demonstrate an amazing visual effect. In this context, virtual view generation technology is beginning to gradually move into the horizon of the public, and is being paid more and more attention in the research of related fields. We can see virtual view generation technology has a board application space and development prospect.There are many solutions for virtual view generation, among these the depth map based virtual view generation algorithm can achieve a fairly good result at a lower cost and ensure real-time. As a consequence this paper select depth map based algorithm to perform the virtual view generation work. In the depth map based virtual view generation algorithm, excepting RGB map, depth map acquired at the same time with RGB map also should be added as reference view. Where the depth map is generated by processing depth data, it can demonstrate the distance from body surface in the scene to the camera in a more intuitive way. There are many methods to obtain depth data, for reasons of efficiency and cost, this paper adopt the method that directly acquire through equipment, and select Kinect 2.0 to obtain depth data. Kinect 2.0 is a somatosensory device released by Microsoft, with the support of the outstanding development kits, Kinect 2.0 is able to obtain higher quality of depth data conveniently and quickly, and it has the advantages that can obtain depth data and RGB data at the same time and has a lower cost.Depth data should convert to depth map for ease of storage, transmission and further processing, the currently common method is linearization method, this paper aiming at the deficiency of this method, proposes cumulative distribution probability based and information entropy based two optimized methods on this basis, and conduct comparative analysis of the depth maps obtained by primary linearization method and two optimized methods through subjective visual effect and objective indicator in different scenes.Optimized depth map should be used as input with the RGB map obtained at the same time to convert to RGBD point cloud data. RGBD point cloud data is the point cloud data which has RGB information, it includes location parameter and RGB parameter. Location parameter is determined by mapping the pixel point in depth map from the two dimensional image coordinate space to the three dimensional world coordinate space, RGB parameter is determined by corresponding RGB map. In order to get exact RGB parameter, this paper proposes depth data based aligning method, completes the calibration of depth map and RGB map, and make the RGBD point cloud data more precise.After obtaining RGBD point cloud data flow, this paper select Direct3 D three-dimensional graph API, input the RGBD point cloud data and vertex coordinatetransformation matrix into GPU, utilizing the parallel computation ability of GPU to complete the generation of virtual view, during this course the shader programming is needed to customize the shader operation so that the arithmetic speed can be increased in order to ensure real-time. The experiment result shows that the virtual view generation algorithm of this paper can real-timely implement the generation of virtual view in a certain angle range, reaches the desired objective of the research.