大工至善|大学至真分享 http://blog.sciencenet.cn/u/lcj2212916

博文

[转载]【计算机科学】【2005.08】图论和热力学框架下三维点云的分割与识别

已有 258 次阅读 2021-6-18 22:16 |系统分类:科研笔记|文章来源:转载

图片


本文为美国东北大学(作者:Anupama Jagannathan)的博士论文,共142页。

 

本文以距离传感器生成的三维点云为特征的自由物体识别为背景,研究了两个基本问题:(1)在输入点云上构造的曲面网格的分割;(2)场景与模型点云的对应关系的确定。许多现有的识别系统需要对模型和场景进行均匀采样,或者假设这些点云重叠。

 

本文描述了在不依赖这些限制性假设情况下的解决方案。网格分割是获得底层对象有效表示的一个重要步骤,并且由于输入数据的噪声而具有挑战性。该方法在输入网格的每个顶点处计算曲率,曲率是一个旋转和平移不变的形状描述符。研究结果表明,该算法与现有的最新方法相比具有很好的鲁棒性,并且在有噪声的情况下具有很好的分割效果。

 

本文的第二个贡献是在识别过程中,当两个点集的基数相差几个数量级时,确定场景和模型点云之间的一一对应关系。图焓和吉布斯自由能的公式捕捉了一对图之间的结构细微差别和基础点集之间的空间差异。通过处理一系列不精确图匹配问题,优化吉布斯自由能,得到所需的对应关系。实验结果表明,该方法在处理干扰和噪声方面优于现有的许多图匹配算法。本文的第三个贡献是减少计算和存储负担,实现实时识别。提出了一种基于图的网格抽取算法,以获得高细节三维曲面网格的保形粗近似。然后导出了一种退化度量,将分层抽取和多尺度对应联系起来。

 

In the context of recognition of free-formobjects, which are characterized by range sensorgenerated 3D point clouds, thisdissertation addresses two fundamental issues (1) segmentation of surfacemeshes constructed over the input point clouds, (2) determination ofcorrespondence between the Scene and Model point clouds. Many existingrecognition systems require uniform sampling of the Model and the Scene or theyassume that these point clouds overlap. This dissertation describes thesolutions to the segmentation and correspondence problems without resorting toany of these restrictive assumptions. Mesh segmentation is an important steptoward deriving an efficient representation of the underlying object and ischallenging due to noisy input data. In the proposed approach, curvedness,which is a rotation and translation invariant shape descriptor, is computed atevery vertex in the input mesh. Iterative graph dilation and morphologicalfiltering of the outlier curvedness values result in multiple, disjoint sub-meshescorresponding to the physical parts of the underlying object. Results indicatethat the algorithm compares well with the existing state-of-the-art approachesand it provides robust segmentations in the presence of noise. The secondcontribution of this thesis is toward the determination of a one-to-onecorrespondence between the Scene and the Model point clouds during recognition,when the cardinalities of the two point sets are orders of magnitude different.Formulations for graph enthalpy and the Gibbs free energy capture thestructural nuances between a pair of graphs and the spatial differences betweenthe underlying point sets. The desired correspondence is obtained by tackling asequence of inexact graph matching problems that optimizes the Gibbs freeenergy. Results indicate that the proposed approach outperforms many existingstate-of-the-art graph matching algorithms in dealing with clutter and noise.The third contribution of this thesis aims at reducing the computational andstorage burden to enable real-time recognition. A graph-based mesh decimationalgorithm is proposed to obtain shape-preserving coarser approximations of ahighly detailed 3D surface mesh. A degradation metric is then derived to linkhierarchical decimation with the multi-scale correspondence.

 

1.  引言

2. 相关工作与技术背景

3. 从非流形到流形表面的转换

4. 面向对象表示的网格分割

5. 图论和热力学框架下的点云匹配

6. 多尺度对应的分层网格抽取

7. 结论与展望


更多精彩文章请关注公众号:205328s611i1aqxbbgxv19.jpg




http://wap.sciencenet.cn/blog-69686-1291762.html

上一篇:[转载]【无人机】【2017.05】基于无人机的大比例尺海岸侵蚀评估制图
下一篇:[转载]【信息技术】【2014.05】用激光传输立体声音频信号

0

该博文允许注册用户评论 请点击登录 评论 (0 个评论)

数据加载中...

Archiver|手机版|科学网 ( 京ICP备07017567号-12 )

GMT+8, 2021-9-27 01:03

Powered by ScienceNet.cn

Copyright © 2007- 中国科学报社

返回顶部