QQ客服 官方微博 反馈留言

题名	单目运动摄像机进行稠密场景重建的方法研究
作者	杨磊
学位类别	博士
答辩日期	2015-05
授予单位	中国科学院大学
导师	李桂菊
关键词	稠密场景模型 单目运动摄像机 联合重建 多序列扩展 两层迭代估计
其他题名	Research on Methods for Dense Scene Reconstruction by Monocular Moving Camera
学位专业	机械电子工程
中文摘要	图形与图像处理的数值方法与硬件设备在过去二十年间都得到了瞩目发展，以场景模型重建作为核心任务的计算机立体视觉领域已逐渐将研究的重点从稀疏场景模型转向稠密场景模型。目前，在单目运动摄像机进行稠密场景重建的研究上, 受制于当前稠密结构重建的初始条件依赖性及算法复杂度等问题,尚难以得到系统、高效且适应性强的方法。 本文研究了对这一方向的部分关键问题，并根据系统方案分为三个主要阶段，分别解决了三个该阶段的主要问题： （1) 针对初始观测阶段，提出了一种两层迭代估计基础矩阵的方法。这一阶段关键问题是建立的两视图间正确的几何对应性。错误的几何对应关系将引发病态的位姿估计及错误的矫正图像，导致初始稀疏重建及稠密立体匹配的错误或失败。通过设置外层向内层传递间接参数的两层迭代模型，结合了迭代线性估计与非线性的七点法，可应对场景结构的平面退化。实验证明，该估计方法性能较为稳定，且具备一定的参数适应性。 （2）针对连续观测阶段，提出了一种多序列扩展式重建方法。这一阶段关键问题是克服景深歧义性导致的系统观测与重建效率下降。短基线系统的景深歧义性在单序列扩展式重建中将不可避免地持续累积，导致观测与重建效率的持续下降直至系统终止。通过建立测度配准模型，采用序列式扩展作为子系统，可进行多路扩展观测并统一配准重建。实验证明：结合在线预处理方法，可提高稀疏扩展式场景重建的效率。 （3）针对稠密重建阶段，提出了一种联合景深与运动重建的方法。这一阶关键问题是克服稠密重建对系统初始条件的依赖性。对初始条件较敏感的稠密重建方法，难以在初始景深估计较差、存在视点变化且帧率较低的单目运动摄像机观测系统下使用。通过建立统一优化场景深度与空间运动的模型，并进行色彩一致性下纹理合成误差检验，可降低重建方法的初值依赖。实验表明：该方法可在一定程度上应对初始条件较差的稠密重建问题，并扩展了运动场重建方法的应用范围。
英文摘要	During the passed two decades, remarkable progress has been made in both numeric methods and hardware devices by related fields of computer graphic and image processing. By considering the scene reconstruction as the core mission, the interest of computer stereo vision community has been gradually shifted from sparse scene model to dense scene model. Currently, on the issue of implementing dense scene reconstruction by using monocular moving camera, systematic, efficient and adaptable method has yet been found, due the there dependency on initial conditions and complexity of the algorithm itself. Parts of the issue had been investigated, and three key-problems respectively from three main stages which could be divided from the solution has been solved. (1) A two-layer iterative estimation for fundamental matrix was proposed on initially observing stage which is in order to solve the key-problem of build correct geometric correspondence between two view. Wicked pose estimation and wrongly rectified images-pair can be triggered by incorrect geometric correspondence. By setting two-layer iterative model for passing the indirect parameters from external layer to inner layer, combined iterative linear estimation with seven points' no-linear method, planar degeneration can be solved. The experiment shows that robust performance of the method can be demonstrated addition with some degree of adaptive parameter. (2) A multiple sequences' solution for extending reconstruction was proposed on continuously observing stage which is in order to solve the problem of undermined system efficiency of observation and reconstruction by depth ambiguity. Depth ambiguity of short baseline system would be incrementing during the extending reconstruction by single sequence which can be considered as unavoidable. The efficiency of observation and reconstruction can be constantly lowing down until the system has to be terminated. By setting metric registration model combined with sequential extending reconstruction as subsystem, multi-way observation and unified reconstruction can be achieved. The experiment shows that the efficiency of sparse extending reconstruction can be improved. (3) A combined reconstruction with depth and motion was proposed on dense reconstruction stage which is in order to solve the problem of dependency on initial conditions in dense reconstruction. The method with high sensitivity on initial value can be hardly implemented under the monocular moving camera observation system which initial depth estimation could be inferior, addition with view changing and low frame-rate. By building an unified model for optimizing scene depth and spatial motion and implementing color-consistency error checking of texture synthesis, the dependency on initial value can be reduced. The experiment shows that the method can be used for implementing dense reconstruction under inferior initial conditions, meanwhile, the application range of motional field's reconstruction has been expanded.
公开日期	2015-12-24
内容类型	学位论文
源URL	[http://ir.ciomp.ac.cn/handle/181722/48938]
专题	长春光学精密机械与物理研究所_中科院长春光机所知识产出
推荐引用方式 GB/T 7714	杨磊. 单目运动摄像机进行稠密场景重建的方法研究[D]. 中国科学院大学. 2015.

个性服务

查看访问统计

相关权益政策

暂无数据

收藏/分享

除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。