【作者】 方磊；

【导师】 王宏远；

【作者基本信息】 华中科技大学 ， 信息与通信工程， 2007， 博士

【摘要】 由图像序列恢复三维场景结构与摄像机运动是计算机视觉领域中重要的应用之一。三维场景重建分为基于特征与基于光流场两种方法,本文主要研究基于特征的图像序列三维场景恢复技术。提出了一种加速鲁棒性参数估计策略:MLESAC-A,通过设置预检验与后检验,不仅可以过滤病态抽样,还能使MLESAC能够采用动态次数抽样的方法从而提高算法效率。合成图像的实验表明加速策略的效率得到极大提高,且当outlier的比例低于30%时,RANSAC和MLESAC-A(EM)以及MLESAC-A(ME)所消耗的时间没有明显区别,RANSAC还略优一些;但当此比例超过30%时,RANSAC耗时呈指数增加,而无论是EM还是ME,MLESAC-A耗时并没有显著增加,表明MLESAC-A比RANSAC更稳定。指出了传统的鲁棒性策略是基于一维数据,这将限制获取的对应点数量从而影响后续的场景重建效果。提出了基于二维数据的MLESAC策略,用每组对应点的匹配点数与匹配强度指导抽样过程。对简单场景的图像对与复杂场景的图像对进行了实验,对应点的数量分别提高了16.7%与56.8%。提出了一种简化的三视图几何约束关系的实现策略,避免了求解三焦张量,通过对三视图进行全局的特征点匹配过程,能够获得最大数量的对应点,并通过误差矩阵对全局匹配进行限制,提高了搜索效率。以三视图为重建单元,提出了一种可并行迭代式分层射影重建策略。使用三视图的几何关系,避免了传统迭代重建中两视图几何关系的不确定性。可并行分层方法不断合并重建单元的对应点,直至最终只剩一个重建单元。对长度为n的序列,传统迭代方法需要进行n-1次重建单元合并,且重建层数为n-1;可并行迭代策略需要进行n-2次单元合并,重建层数为[log 2( n? 1)],且每一层的单元合并完全可并行,从而能够提高重建过程的效率。为了避免迭代式重建策略所存在的累积误差效应,提出了一种线性回溯射影结构恢复策略。每恢复一幅图像之后,将所有当前可见的结构用最新的数据进行回溯估计,并用新的结构更新摄像机的运动。回溯步骤会降低重建效率,然而由于重建方法全部是线性算法,而且结构与运动的恢复过程相比对应点合并过程,其增加的计算量可以忽略,因此对重建过程的效率影响非常小。对由上述方法获得的射影重建的初始结果利用非线性优化方法——光束法平差进行了非线性优化。利用Pollefeys的线性自标定方法快速的将射影重建的结果恢复至度量空间,在度量空间再次使用光束法平差进行非线性优化从而获得最终的在度量空间下可视化的场景结构与摄像机运动。更多还原

【Abstract】 The recovery of structures of the scene and motion of the cameras from image sequence is one of the most important applications in computer vision. There are two kinds of methods of the 3d scene reconstruction: one is based on features, the other is based on optical flows. This paper researches the technology of 3d scene reconstruction from image sequence based on feature.An accelerated robust parameter estimation strategy named MLESAC-A is presented. By set preview and post verification, not only can it filters degenerate samples, but also can it adopts dynamic sample method so that it can improve the speed of the algorithm. Experiment on synthesis images show that MLESAC-A is much faster than MLESAC, and when the proportion of outliers is lower than 30% the consumed time of RANSAC, MLESAC-A(EM) and MLESAC-A(ME) is almost the same, but when the proportion is higher than 30% the time of RANSAC increases exponentially while the time of MLESAC-A(EM) and MLESAC-(ME) increase indistinctively, this shows MLESAC-A is more stable than RANSAC.A conclusion that traditional robust strategy like RANSAC is based on 1D data is pointed out, this will restrict the number of correspondence so that the quality of reconstruction will be reduced. A MLESAC strategy based on 2D data is advanced, it uses matched number of each group of correspondence and matched score to guide sample process. Experiments on simple scene and complex scene show that the number of correspondence has been increased by 16.7% and 56.8% respectively.A simplified method that implements triple-view geometry constraint is put forward to substitute for the computation of trifocal tensor. By global features matching among three views, it can obtains maximum number of correspondence, and by the limitation of error matrix, the efficiency of the global matching process will be improved.A parallel iterative hierarchical projective reconstruction strategy is presented, which is organized on the reconstruction unit of triple-view. The usage of three view geometry constraint can avoid the uncertainty in epipolar geometry constraint which is used in traditional iterative reconstruction. The parallel method combines continues reconstruction unit until there is only one unit left. For a sequence of size n, the traditional method needs n-1 combinations and n-1 layers; while the parallel method needs n-2 combinations and [log 2( n ? 1)] layers, in each layer the combinations are parallel, which can improve efficiency of the reconstruction.To avoid the effect of cumulate error in conventional iterative reconstruction algorithms, a linear rewound method is put forward. In each iterative step, after the reconstruction of the current view, use the information to re-estimate all the structures that can be seen then, and update the motion of the cameras. The rewound step will reduce the efficiency of the reconstruction process, since all the methods used in reconstruction are linear and the incremental computation of structure from motion process can be ignored compared to the combinations of correspondence process, it will affect the reconstruction very slightly.A nonlinear optimization algorithm, bundle adjustment, is used to optimize the initialize projective reconstruction result obtained by above methods. A Pollefeys’s linear self-calibration method is adopted to upgrade the result from projective space to metric space, bundle adjustment is again used in metric space to achieve a final visible result of structures of the scene and motion of the cameras.更多还原