大场景分段曲面重建和优化方法

发布时间：2018-06-30 05:16

本文选题：表面重建 + 网格优化　；参考：《浙江大学》2017年硕士论文

【摘要】：从输入点云中提取高质量的表面在计算机图形图像中是很重要的一个研究方向。主要内容包括了如下几个方面:对已有点云的拟合,对表面漏洞的填充以及对已有模型的重构。在本文中,我们首先从初始的采样点中通过泊松碟采样提取出顶点集。接着利用顶点集做Voronoi划分得到初始点云的邻接关系并从中提取出粗糙的初始网格。在Voronoi划分中,我们采用了最短路径作为距离函数,这样可以在一定程度上降低初始点云的不均匀分布对实验结果的影响。在获得了初始网格之后,我们可以发现初始网格的拓扑结构和几何信息都存在明显的不足,因此需要选择合适的方法来优化初始网格。我们采用了将几何信息和拓扑信息作为一个整体进行优化的框架。整个优化过程通过顶点集V和邻接关系矩阵B被表示为带约束的q范数优化问题。最后我们通过最小化由点云到三角网格的距离和正则项组成的能量函数就可以找到最合适的顶点位置和三角网格的拓扑结构。在我们的优化算法中,法向信息并不是必需的。现在的许多表面重建的算法都会要求我们提供法向信息。众所周知,法向评估对于误差和尖锐特征非常敏感,这也导致在面对法向质量不高的点云时,许多算法的结果会有较大的误差。而我们的算法在面对这样的数据时依然可以获得较好的效果。虽然我们在初始化过程中需要用顶点法向来判断三角面片的朝向,但对法向的精度要求并不高。由于本文所提出的算法在更新顶点几何信息和拓扑结构时只需要顶点附近的采样点的几何信息,我们可以将初始点云分割成多个子块,对每一个子块依次重建出合适的表面,最后通过拼接的方法得到整个场景的重建结果。实验表明,在处理一些较大规模的点云时,这种生成表面的方式在运行效率和生成的网格质量这两方面的表现都很出色。
[Abstract]:The extraction of high quality surfaces from the input cloud is an important research direction in computer graphics and images. The main contents include the following aspects: the fitting of the existing cloud, the filling of the surface holes and the reconstruction of the existing models. In this paper, we first extract the vertex set from the initial sampling point by Poisson disc sampling. Then the vertex set is used as Voronoi partition to obtain the adjacency of the initial point cloud and the rough initial grid is extracted from it. In Voronoi partition, we use the shortest path as the distance function, which can reduce the influence of the uneven distribution of the initial point cloud on the experimental results to some extent. After obtaining the initial grid, we can find that the topology and geometric information of the initial grid are obviously inadequate, so we need to select a suitable method to optimize the initial grid. We adopt a framework that optimizes geometric and topological information as a whole. The whole optimization process is represented by vertex set V and adjacency matrix B as Q norm optimization problem with constraints. Finally, by minimizing the distance from the point cloud to the triangular mesh and the regular energy function, we can find the most suitable vertex position and the topological structure of the triangular mesh. In our optimization algorithm, normal information is not necessary. Many surface reconstruction algorithms now require us to provide normal information. It is well known that normal evaluation is very sensitive to errors and sharp features, which also leads to large errors in the results of many algorithms in the face of point clouds with low normal quality. However, our algorithm can still get good results in the face of this kind of data. Although we need to use vertex method to judge the orientation of triangulation, the accuracy of normal direction is not high. Since the algorithm proposed in this paper only needs the geometric information of the sample points near the vertices when updating the vertex geometry information and topology structure, we can divide the initial point cloud into several sub-blocks and reconstruct the appropriate surface for each sub-block in turn. Finally, the reconstruction result of the whole scene is obtained by splicing. The experimental results show that this surface generation method performs well both in terms of running efficiency and mesh quality when dealing with some large scale point clouds.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】