基于视频的三维重建研究

发布时间：2018-01-28 19:53

本文关键词： 三维重建单目视频双目视频视差场景流　出处：《华中科技大学》2016年硕士论文　论文类型：学位论文

【摘要】：基于视频的重建技术在计算机视觉领域中扮演着非常重要的角色,而如何恢复场景的三维模型是目前研究的热点与难点问题。本文围绕基于视频的场景重建技术展开讨论,包括基于单目视频的三维场景重建和基于双目视频的视差图和场景流获取。由于单目包含的深度信息比较少,如何基于单目视频恢复相机的运动参数以及目标的深度信息是研究的重点与难点。双目视频虽然包含了非常显著的深度信息,但是考虑到视频中场景的连续性问题,如何使得恢复出的深度图保持前后帧的连续性以及场景中运动目标的一致性,也是比较困难的问题。因此,针对上述所提到的问题进行了深入的研究,具体的研究工作如下:第一,对三维重建研究进行了详细的介绍,介绍了对于特征点匹配的理解以及我们提出的基于特征引导偏向性高斯混合模型(Feature Guided Biased Gaussian Mixture model,FGBG);详细介绍立体视觉中立体匹配算法的原理、分类及评测标准,并在4个典型的数据集上对有代表性的局部、全局、半全局算法进行对比实验。此外,详细介绍运动恢复结构(SFM)的基本原理,并进行了实验分析。第二,提出一种基于双目视频的视差图和场景流获取技术。基于双目视频,首先获得初始的视差图和2D特征点轨迹;在此基础上获得初始的3D稀疏运动轨迹,利用本文提出的Object Motion Hypothesis(OMH)算法获得运动物体的一致性假设;采用slanted-plane model以及参考图像与前后时间点图像对的约束关系,构建超像素和运动物体之间的能量模型,通过优化获得视差和场景流的估计结果。第三,提出一种基于单目视频的动态场景重建系统。在获取特征点轨迹的基础上,基于运动信息获得特征点轨迹的聚类结果;提出一种基于超像素的多标记Graph-cut算法,得到每一个目标的精确边界;为每一个运动目标分配一个虚拟相机;通过标准的SFM方法分别单独估计每个运动目标对应的虚拟相机的参数和稀疏三维点云,通过PMVS和泊松表面重建获得目标的稠密重建结果。
[Abstract]:Video based reconstruction technology plays a very important role in the field of computer vision. How to restore the 3D model of scene is a hot and difficult problem. This paper focuses on the video based scene reconstruction technology. It includes 3D scene reconstruction based on monocular video and parallax map and scene stream acquisition based on binocular video. How to recover camera motion parameters and depth information based on monocular video is the key and difficulty of the research. Although binocular video contains very significant depth information. However, considering the continuity of the scene in the video, how to make the restored depth map maintain the continuity of the frame before and after and the consistency of the moving object in the scene is also a more difficult problem. The research work is as follows: first, the 3D reconstruction research is introduced in detail. This paper introduces the understanding of feature point matching and our hybrid model of feature-based guided bias Gao Si (. Feature Guided Biased Gaussian Mixture model. FGBGN; The principle, classification and evaluation criteria of stereo matching algorithm in stereo vision are introduced in detail, and the typical local, global and semi-global algorithms are compared with each other on four typical data sets. The basic principle of motion recovery structure (SFM) is introduced in detail, and the experimental analysis is carried out. Secondly, a parallax image and scene stream acquisition technology based on binocular video is proposed, which is based on binocular video. First, the initial parallax map and 2D feature point locus are obtained. On this basis, the initial 3D sparse motion trajectory is obtained, and the consistency hypothesis of moving objects is obtained by using the Object Motion HypothesisOMH algorithm proposed in this paper. The energy model between superpixel and moving object is constructed by using slanted-plane model and the constraint relation between reference image and image pair. The estimation results of parallax and scene flow are obtained by optimization. Thirdly, a dynamic scene reconstruction system based on monocular video is proposed. The clustering results of feature point locus are obtained based on motion information. A multi-label Graph-cut algorithm based on super-pixel is proposed to obtain the exact boundary of each target. Assign a virtual camera for each moving target; The parameters and sparse 3D point clouds of each moving target are estimated separately by the standard SFM method. The dense reconstruction results of the target are obtained by PMVS and Poisson surface reconstruction.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】