基于Kinect的抠像算法研究与应用

发布时间：2018-06-10 14:06

本文选题：Kinect + 深度信息　；参考：《北京邮电大学》2016年硕士论文

【摘要】：随着多媒体技术的发展和数字传媒业的科技化,数字抠像和合成作为图像处理领域炙手可热的研究课题,在电影、游戏和广告制作业,以及医疗卫生和空间探测等多领域都有广泛的应用。数字抠像技术是指通过算法在图像和视频中提取带有透明度的感兴趣前景。传统的自然抠像算法大都需要用户提供划分好的三元图或涂鸦信息作为提示信息,耗费大量的时间和人力成本,且很难应用于视频抠像。随着深度信息获取技术的发展,一系列结合深度信息的RGBD抠像算法应运而出。本文主要研究基于Kinect深度信息的自然图像抠像算法,最终实现无需人工输入辅助信息的全自动抠像系统。论文主要完成工作如下:(1)研究了体感外设Kinect的起源、发展和应用,比较并总结了一代和二代传感器的硬件配置和深度图像获取技术。针对实验中Kinect二代的深度图像误差,将结合深度信息的RGB-D引导滤波引入到深度图像的修复中,并将其进行迭代计算以提高对深度图边缘和内部空洞的修复精度。(2)分析及对比了几种传统自然图像抠像技术的原理和应用场景,并深入研究了基于TOF的深度抠像算法,将Kinect获取的深度信息引入传统的RGB抠像算法中。综合考虑抠像的实时性和精度,选择了 Shared Matting作为本文研究算法。(3)实现了三元图的自动生成算法,并将深度信息引入Shared Matting,提出了基于深度域改进的区域扩张方法。最终实现了基于Kinect深度信息的全自动抠像算法,经测试本算法具有较好的鲁棒性和实时性。(4)设计并搭建了带有可视化界面的Kinect实时抠像和合成系统。针对抠像算法中的人物前景与新的虚拟背景之间存在的色调和光照条件的不一致,在系统中引入保持色调一致性的Erik Reinhard色调迁移算法。最终实现了 Kinect彩色视频流的实时人体抠像和合成,并能对前后景色调一致的新的“人景合一”的图像进行存储。
[Abstract]:With the development of multimedia technology and the technology of digital media industry, digital matting and synthesis as a hot research topic in the field of image processing, in film, games and advertising industry, As well as medical and health and space exploration and other fields have a wide range of applications. Digital matting technology is used to extract the interesting foreground with transparency in image and video. Most of the traditional natural matting algorithms require users to provide divided ternary or graffiti information as prompt information, which cost a lot of time and manpower, and it is difficult to be applied to video matting. With the development of depth information acquisition technology, a series of RGBD matting algorithms combined with depth information should be carried out. This paper mainly studies the natural image matting algorithm based on Kinect depth information, and finally realizes the automatic matting system without manual input of auxiliary information. The main work of this paper is as follows: (1) the origin, development and application of Kinect are studied, and the hardware configuration and depth image acquisition techniques of generation and generation 2 sensors are compared and summarized. Aiming at the error of Kinect second generation depth image, the RGB-D guided filter combined with depth information is introduced into the depth image restoration. In order to improve the repairing accuracy of depth map edge and inner cavity, the principles and application scenes of several traditional natural image matting techniques are analyzed and compared, and the depth matting algorithm based on TOF is deeply studied. The depth information obtained by Kinect is introduced into the traditional RGB matting algorithm. Considering the real-time and precision of matting, this paper chooses shared matching as the algorithm to realize the automatic generation of ternary images, and introduces the depth information into shared tracking, and proposes an improved region expansion method based on depth domain. Finally, the automatic matting algorithm based on Kinect depth information is realized. The Kinect real-time matting and synthesis system with visual interface is designed and built by testing the algorithm has good robustness and real-time performance. Aiming at the inconsistency of hue and illumination conditions between the foreground of characters in matting algorithm and the new virtual background, Erik Reinhard hue migration algorithm is introduced in the system. Finally, the real-time human matting and synthesis of Kinect color video stream is realized, and the new "human scene in one" image which is consistent with the front and rear scenery can be stored.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】