基于GPU的高清视频虚拟融合技术研究

发布时间：2018-03-13 16:04

本文选题：虚拟广告　切入点：GPU　出处：《杭州电子科技大学》2013年硕士论文　论文类型：学位论文

【摘要】：随着电子产业的发展、硬件成本下降以及图像视频处理各种算法不断推陈出新，视频处理技术在工业、商业以及民用方面已经达到实用的阶段。在视频处理技术快速发展的环境下，，虚拟广告系统应运而生。虚拟广告技术在电视节目播出中提供大量虚拟信息，如三维动画、虚拟广告牌、虚拟信息提示等。这些虚拟信息提供大量观众单纯从节目本身无法或很难得到的信息，不仅丰富了节目的内容，而且还提供了新的广告形式，能产生巨大的经济效益和社会效益。虚拟融合技术使得虚拟物体与视频场景在几何位置和光色度上保持一致性，虚拟物体完全融入到视频场景中，给观众强烈的真实感。本文主要对高清体育视频中实时摄像机运动跟踪技术和虚拟广告区域快速前景分割技术进行了研究，旨在研究虚拟融合技术在高清数字视频时代的可行性和实用性。首先，设计并实现了高清音视频同步采集系统，实时采集HD-SDI高清音视频数据。本文采用DirectShow来实现高清音视频的采集功能。高清音视频同步采集系统可以针对不同的视频采集卡进行高清音视频采集，并实现了Filter过滤器中的回调函数，可以在回调函数中实现对视频的融合以及为其他模块处理提供视频流数据。其次，利用GPU中CUDA框架的并行运算和高速浮点计算特性，提出并实现了基于GPU的摄像机运动跟踪算法。论文阐述了基于自然特征点匹配的运动跟踪算法和基于Hough空间变换的运动跟踪算法的具体实现流程。通过GPU加速和优化的摄像机运动跟踪算法，将求解场地线端点信息和获得摄像机的投影变换矩阵的时间消耗，从每帧193ms提升到了每帧16ms，大大提高了摄像机跟踪算法在处理高清视频的能力。最后，提出了基于可靠背景模型的运动目标分割算法和基于GPU的运动目标分割算法。基于可靠背景模型的运动目标分割算法能够达到了准确分割前景目标的要求，但在实际应用中却发现存在相机大幅度移动和高清视频大数据实时处理问题。基于GPU的运动目标分割算法，主要利用GPU并行处理技术来动态构建背景模型和分割运动目标。基于GPU的运动目标分割算法技术很好地解决了摄像机移动和高清视频大数据量实时处理的问题。本文提出的高清视虚拟融合技术的可行性在高清音视频采集融合系统中得到充分验证。系统表明这几项技术在不同体育视频中具有较好的实用效果。
[Abstract]:With the development of electronic industry, hardware cost reduction and a variety of video processing algorithms constantly, video processing technology in the industrial, commercial and civil areas has reached the practical stage. In the environment of the rapid development of video processing technology, virtual advertising system came into being. Virtual advertising in the TV broadcast to provide a large number of virtual information, such as 3D animation, virtual billboards, virtual information. These virtual information to provide a large number of viewers simply from the program itself unable or difficult to get information, not only enrich the content of the program, but also to provide a new form of advertising, can produce enormous economic benefits and social benefits. The virtual fusiontechnology makes virtual objects and video scene to maintain consistency in geometry and light, virtual objects are fully integrated into the video scene, give the audience a strong sense of reality. In this paper, the real-time camera motion tracking technology and the fast foreground segmentation technology of virtual advertising area are studied, aiming at studying the feasibility and practicability of virtual fusion technology in the era of high-definition digital video.
First of all, the design and implementation of synchronous acquisition system of high-definition video, real-time HD-SDI HD audio and video data. This paper uses DirectShow to achieve high-definition audio and video capture functions. HD audio and video synchronization acquisition system for different video capture card for high-definition audio and video acquisition, and Filter callback function in the filter, can be realized the video fusion and provide video data for other modules in the callback function.
Secondly, based on the characteristics of computing parallel computing and high-speed floating-point CUDA GPU framework, proposed and implemented cameramotion tracking algorithm based on GPU. This paper expounds the motion tracking algorithm based on natural feature point matching and tracking algorithm in Hough space transform based motion tracking algorithm. Through the specific implementation process of GPU acceleration and optimization of camera motion that time will solve the field line endpoint information and projection transform matrix of camera consumption increased from 193ms per frame to 16ms per frame, greatly improving the camera tracking algorithm in the ability to handle HD video.
Finally, this paper proposed a moving object segmentation algorithm and reliable background model GPU based moving object segmentation algorithm based on background model. Reliable moving target segmentation algorithm to achieve accurate segmentation of foreground objects based on the requirements, but in practice there exist the problem of real-time processing and camera greatly mobile HD video data segmentation algorithm. The moving target based on GPU, the main use of GPU parallel construct background model and segmentation of moving objects in dynamic processing technology. GPU moving target segmentation algorithm solves the mobile camera and HD video data real-time processing based on the problem.
The feasibility of the HDV virtual fusion technology presented in this paper is fully verified in HD audio and video acquisition and fusion system. The system shows that these technologies have good practical effects in different sports videos.

【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】