【作者】 刘宇；

【导师】 鲍虎军； 章国锋；

【作者基本信息】 浙江大学 ， 计算机应用技术， 2016， 硕士

【摘要】 运动捕捉(Motion Capture)是计算机视觉领域的基础研究课题之一,旨在基于多个不同视角的同步相机阵列恢复人体运动过程中的三维模型和细节特征,在虚拟现实(Virtual Reality)、数字动画、影视制作等领域有着广泛应用。当前主流的基于同步相机阵列的运动恢复虽然在三维模型重建方面取得了一些较好的成果,但主要存在三方面的问题:普通设备无法进行多个视角连续高速捕获的数据获取需求;不能够实现自动分割,需要手动抠取大量的Mask作为输入;往往受制于内存大小限制而无法获得高精度的重建效果。本文基于浙江大学人体反射场高性能捕获系统,通过其同步高速采集的视频序列,进行运动人体三维模型重建。首先,我们设计并搭建了一套完备的基于同步摄像机阵列的自动人体三维重建系统,涵盖了数据采集和标定、自动分割、构建VisualHull可见外壳、深度恢复、点云采样和三维模型重建等模块。其次,提出了一个结合颜色、纹理、对比度和边缘等多线索的自动前景分割方法:先在灰度空间利用背景差分进行前景初始化,其次用Graph Cuts得到初步的分割结果,最后应用Background Cut降低背景区域的对比度,将三种方法的优点很好地结合起来。实验结果表明,我们设计搭建的软件系统合理有效,不仅能够恢复出高精度的深度结果和获得密度可控的点云采样,而且能够稳定地重建出细节突出、表面平滑的人体运动三维模型。所提出的多线索分割算法,相比其他分割方法,能够有效降低光照和前背景颜色相似对分割的影响,实现了高质量的前景自动分割,为自动的人体三维模型重建奠定了基础。更多还原

【Abstract】 Motion Capture is one of the basic research fields in computer vision,which aims to recover the three-dimensional models and details of texture from moving people,which are captured by a bunch of synchronized cameras from different views.Motion Capture has very broad applications ranging from Virtual Reality,Digital Animation to Movie Making.The emergence of many excellent works based on synchronized camera array has achieved a series of good results in 3D reconstruction,but they are always accompanied with some limitations and problems in three aspects:Firstly,original devices could not satisfy the needs of capturing successive data from different views with high speed;Secondly,automatical segmentation could not be implemented,which makes it troublesome for users to cutout massive images manually.Finally,the results of reconstruction are often limited by the memory thus could not reach to a relatively high-resolution level.This paper,based on the Reflection Field of Human Body Capture Platform in Zhejiang University,with which we can capture synchronized video sequences with high speed and conduct three-dimensional reconstruction of moving people.Firstly,we design and build a complete system to recover human body based on synchronized camera array,which contains modules of data acquisition and calibration,automatical segmentation,visuallhull construction,depth recovery,pointcloud sampling and 3D reconstruction.We also propose a novel segmentation method which combines multiple cues such as color,texture,contrast,edge to cutout foreground automatically:the first step is to initialize foreground with background subtraction in gray space,then followed with an initial segmentation using graph cuts,and background cut is implemented in order to attenuate the background contracts to remove wrong segmented results.This algorithm combines the merits of three methods effectively.The experimental results show that,on the one hand,our software is useful and efficient,which can not only recover depth map with high-resolution as well as get density controllable pointCloud sampling,but obtain moving people’s 3D models with texture details and smooth surface.On the other hand,compared with other segmentation methods,our algorithm could effectively reduce the impact for segmentation brought by light and color similarity between foreground and background,and the high quality foreground cutout results lay the foundation for automatically three-dimensional of human body.更多还原