| With the rapid development of global science and technology,twodimensional display has been unable to meet the requirements of users for visual scenes in many application fields.Three-dimensional visualization is widely used in games,movies,architectural design,medical treatment,military and other fields.The true three-dimensional light field display technology with the characteristics of large field of view,high resolution and true color is known as "the most vital next-generation display technology".It is an important development direction in the field of opticelectronic information.With the upgrade of hardware and the advancement of network technology,the problems of data calculation and data transmission involved in 3D light field display have been gradually solved,but the 3D data acquisition and processing technology of target objects in dynamic light field still has limitations.The problems of extracting objects from images and reducing the complexity of data processing need to be solved urgently.This thesis conducts research on the collection and processing of human body information under 3D dynamic light field,and the research results are summarized as follows:Based on the existing portrait matting algorithm,a new multi-view portrait matting algorithm based on deep learning is designed by taking advantage of the characteristics of regular arrangement and no occlusion relationship of the multi-view images captured by the acquisition platform.The algorithm can perform portrait matting on multiple images at the same time.Although its matting accuracy has declined,its matting speed is about 40 times higher than that of the existing algorithm,which is more suitable for the real-time data processing of light field video live broadcast.This method abandons the conventional idea of improving the matting speed of a single image,and combines the actual task scene to improve the matting speed as a whole.The algorithm requires a pre-shot background image without people to assist the matting network in matting.For the images with different background complexities and different number of viewpoints,it can maintain a relatively stable matting result.A light field information acquisition platform with an arc-shaped camera array as the main body is designed and built.And its acquisition angle can reach about 90 degrees.In conjunction with software such as open broadcaster software and hardware such as Multi-view,3D dynamic light field portrait capture with 48 viewpoints is realized.After the collected 3D human body data is transmitted through the network,a large amount of data processing is needed to restore the 3D information of the human body.A rendering platform with a 3D video player as the main body is designed and built to complete the data processing tasks such as multiview portrait alignment and multi-view light field encoding.A test system of information acquisition and processing in humanoriented 3D dynamic light field is built.The system integrates 3D portrait acquisition,multi-view video transmission and 3D video rendering.And in the actual test,a 3D light field video live broadcast with a resolution of nearly 8K is realized.With the 3D light field display device independently developed by our team,the target human body has been reproduced in three dimensions,and it has a good visual perception.The ultimate goal of display technology is to reproduce the scenes of the objective three-dimensional world,so that people can obtain the same visual perception as they are in the real world.The research and results of this thesis on the acquisition and processing of 3D dynamic light field information provide new technical support for the light field video live broadcast. |
PDF Full Text Request