基于泊松分布的点云数据栅格化算法研究

发布时间：2018-06-01 08:20

本文选题：Delaunay + 栅格　；参考：《西北农林科技大学》2017年硕士论文

【摘要】：随着对计算机图形学领域研究的进一步深入,其所衍生出的虚拟现实以及增强现实等研究方向吸引了众多的研究学者。同时也对计算机图形学的发展提出了更大的需求,相关产业所涉及的快速建模也一度成为研究学者所热衷的研究方向。点云数据的获取主要通过扫描设备或者从图片中提取特征点,点云数据获取方式的不同所得到的点云密度也不尽相同,研究者针对不同密度点云数据提出了基于稀疏或稠密点云数据的网格重建方法,现有的方法对不同密度点云数据鲁棒性不强,本研究针对点云分布密度不均匀以及无固定的点云密度界定等问题提出了一种提高三维重建普适性的基于点云泊松分布的自适应三维重建方法。本研究基于上述说明问题,提出了改进方法。主要完成了以下工作:(1)获取到的点云数据集呈离散分布,首先对点云数据进行预处理操作和数据分析,为下一步栅格化做论证。进而根据栅格中进行曲面拟合所需最小三维点数及泊松分布适用条件确定栅格阈值范围,然后在阈值范围内动态调整阈值大小并迭代栅格划分步骤,直到阈值可以使得栅格划分单元中点云分布符合泊松分布,从而保证算法处理数据的稳定性和自适应特性的鲁棒性。(2)对确定阈值大小后的点云数据进行栅格划分并建立相应的邻域关系与拓扑关系。在不考虑点云密度的情况下,通过迭代过程中动态调整后的阈值对三维数据点进行空间划分,使得划分之后的栅格单元中点云分布趋于稳定;然后对划分得到的栅格单元中的三维数据点构建邻域的拓扑邻近关系,同时构建栅格单元的26邻域关系,从而保证三角剖分生成网格过程中搜寻扩展点线的范围,提高扩展搜寻速度。(3)选取最佳候选点不断扩展生成三角网格。以栅格划分和建立的拓扑关系为基础,选择合适的初始三角形作为种子三角形遍历点云数据集进行扩展,通过对栅格中点的邻域搜索和对候选点所在栅格邻域范围的搜索查找,直到所有的三维点和生成的线及三角面片都加入网格中,以此来验证自适应栅格划分的效果和算法的鲁棒性。实验结果表明,对比分析三角面片数和运行耗时两个参数,本研究方法在三角剖分面片数增加了2.5%~27.9%,在运行耗时降低了5.0%~31.6%。实验表明在大部分情况下本文算法在运行效率和三角形面片的交叉和误配方面鲁棒性较强。
[Abstract]:With the further development of computer graphics, the virtual reality and augmented reality have attracted a lot of researchers. At the same time, the development of computer graphics has put forward a greater demand, and the rapid modeling involved in related industries has once become a hot research direction for researchers. The point cloud data is obtained mainly by scanning equipment or extracting feature points from the image, and the point cloud density is different from the different acquisition methods of point cloud data. Based on sparse or dense point cloud data, researchers proposed a mesh reconstruction method for different density point cloud data. The existing methods are not robust to point cloud data with different densities. In this paper, an adaptive 3D reconstruction method based on Poisson distribution of point cloud is proposed to improve the universality of 3D reconstruction. Based on the above problems, an improved method is proposed. Firstly, the point cloud data is preprocessed and analyzed to prove the rasterization of the next step. Then the threshold range is determined according to the minimum number of 3D points needed for surface fitting and the applicable conditions of Poisson distribution in the grid, and then the threshold value is dynamically adjusted and the grid partition steps are iterated within the threshold range. Until the threshold value enables the point cloud distribution in the grid partition unit to conform to the Poisson distribution, Therefore, the stability of the data and the robustness of the adaptive property of the algorithm are guaranteed. The point cloud data after determining the threshold value are raster partitioned and the corresponding neighborhood and topology relations are established. Without considering the point cloud density, the point cloud distribution in the grid cell is stable by dividing the 3D data points by the dynamically adjusted threshold in the iterative process. Then, the topological proximity of the three dimensional data points in the grid cells is constructed, and the 26 neighborhood relations of the grid cells are constructed, so as to ensure the range of the extended point lines in the process of triangulation generation. Improve the speed of extended search. 3) select the best candidate point and expand to generate triangular mesh. Based on the topological relation of grid partition and establishment, the suitable initial triangle is selected as the seed triangle to traverse the point cloud data set. By searching the neighborhood of the middle point of the grid and searching the range of the candidate point in the grid neighborhood, Until all the 3D points and generated lines and triangles are added to the mesh to verify the effectiveness of the adaptive grid partition and the robustness of the algorithm. The experimental results show that the number of triangles and the running time are compared and analyzed. The number of triangulated slices is increased by 2.5% and 27.9%, and the running time is reduced by 5.0% and 31.6%. Experiments show that in most cases, the proposed algorithm is robust in terms of running efficiency and cross and mismatch of triangular surfaces.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】