【作者】 孙小平；

【导师】 佟国峰；

【作者基本信息】 东北大学 ， 模式识别与智能系统， 2015， 硕士

【摘要】 大规模场景点云的表面重建主要针对激光扫描或者图像重建后的点云数据进行网格化重建。表面重建算法的研究在遗产保护,城市建模,虚拟视觉等领域有十分重要的应用价值。本文主要针对序列图像重建后的大规模场景点云数据进行表面重建算法的研究,对两种不同的表面重建算法进行了研究与改进。同时提出了基于GPU的并行化泊松表面重建算法,在保证精度的同时提高了算法的快速性。首先,本文利用序列图像进行大规模场景的三维重建。通过序列图像进行特征提取与匹配,利用多视图几何理论恢复场景的三维结构和摄像机的位置姿态,并利用捆集调整算法进行约束提高重建的精度,由于重建是基于图像特征提取的,所以获得的点云数据十分稀疏,不能恢复场景的精细结构,因此本文利用基于面片的多视图立体匹配算法进行稠密重建,得到了大规模场景的精细结构,对重建后的点云进行表面重建。然后,研究了两种经典的表面重建算法来进行大规模场景的点云表面重建。本文首先研究了基于Delaunay三角剖分的经典算法Power Crust表面重建算法,该算法利用中心轴变换来进行表面重建,得到了平滑的表面模型,但是Delaunay三角剖分的计算复杂,重建时间较长,因此本文提出了降采样的改进算法。然后,针对基于隐函数的泊松表面重建算法进行了研究与改进,针对泊松表面重建的过度平滑的问题,本文提出了加入点集约束的表面重建算法,并将问题转化为求解屏蔽的泊松方程的过程,改进后的算法提高了重建的精度。之后,本文重点介绍了基于GPU的并行化表面重建算法。首先根据给定的输入点集创建八叉树,然后在八叉树空间上计算隐函数,最后利用移动立方体方法提取等值面获得点云的三角网格模型。该算法在GPU上实时创建八叉树,并创建了两个查找表来提高邻近信息的访问速度,为表面重建中的隐函数计算和等值面提取提供了方便。本文在GPU上并行求解泊松方程和提取等值面,从而实现了整个算法的并行化。最后,本文提出的泊松表面重建改进算法可以得到更加逼近输入点集的三角网格,提高了表面重建的精度。同时并行化的重建提高了算法的速度,本文算法在提高精度和速度方面都做出了重要改进,从而得到了大规模场景点云的表面模型。更多还原

【Abstract】 The Surface reconstruction of large-scale scene point cloud is mainly focused on laser scanning point cloud data or point cloud which was reconstructed from sequence image,it has an important role in heritage conservation,urban modeling,virtual vision and other fields.This thesis mainly research Surface reconstruction algorithm for large-scale scene point cloud data sequence reconstructed from image,and Poisson surface reconstruction algorithm is applied to the scene point cloud surface reconstruction.This thesis present a Poisson surface reconstruction improved algorithm that introduces interpolation constraint.This paper also proposed Poisson surface reconstruction based GPU parallelization for speedability while ensure accuracy.Firstly,large-scale 3D reconstruction from image sequences scene.Through image sequences SIFT feature extraction and matching,using multi-view geometry theory recovery camera position and attitude of the three-dimensional structure of scene.Then use the bundle adjustment algorithm set constraints for improving the accuracy of the reconstruction,because reconstruction is based on image feature extraction,so the point cloud data are very sparse and can not recover the fine structure of the scene,so this thesis reconstructs dense point cloud using multi-view matching algorithm(PMVS),then obtain fine scene structure for surface reconstruction.Secondly,this thesis focuses on two classical surface reconstruction algorithm for large-scale scene point cloud surface reconstruction.This thesis is mainly focused on The PowerCrust surface reconstruction algorithm and Poisson surface reconstruction algorithm.The Power Crust surface reconstruction algorithm use a central axis conversion to reconstruct surface based on Delaunay Triangulation.This reconstruction algorithm obtain a smooth surface model,but Delaunay Triangulation processes are complex and take a lot of time,So this thesis puts forward the improved algorithm based on downsampling.Poisson surface reconstruction algorithm is a implicit function algorithm which can obtain a smooth surface,but it has the smoothing over problems,so this thesis propose a interpolation constraint reconstruction algorithms,The extension can be interpreted as a generalization of the underlying mathematical framework to a screened Poisson equation.The improved algorithm can obtain higher-quality surface.Then this thesis presents a parallel surface reconstruction algorithm that runs entirely on the GPU.Like existing implicit surface reconstruction methods,the algorithm first builds an octree for the given set of oriented points,then computes an implicit function over the space of the octree,and finally extracts an isosurface as a water-tight triangle mesh.A key component of our algorithm is a novel technique for octree construction on the GPU.This technique builds octrees in real-time and uses level-order traversals to exploit the fine-grained parallelism of the GPU.Moreover,the technique produces octrees that provide fast access to the neighborhood information of each octree node,which is critical for fast GPU surface reconstruction.GPU algorithm performs Poisson surface reconstruction,which produces high quality surfaces through a global optimization.Finally,the proposed Poisson surface reconstruction algorithm can get fine structure the scene,at the same time,the parallel reconstruction algorithm improves the speed of reconstruction for large-scale scene point cloud.The final experiment proved that the accuracy of the algorithm didn’t change and the parallel algorithm greatly improved the algorithm’s speed.更多还原