１
目录
引言
1．选题背景 ３
1.1 问题的提出 ３
1.2 国内外研究现状 ４
1.2.1 国内研究现状 ４
1.2.2 国外研究现状 ４
1.3 研究的目标、内容及技术路线 ５
1.3.1 研究目标 ５
1.3.2 研究内容 ５
1.3.3 技术路线 ６
2 稀疏三维点云的生成 ６
2.1 图像序列的获取 ６
2.2 相机标定的计算与存储 ７
2.3 特征点检测与匹配算法 ８
2.3.1 特征点检测 ８
2.3.1.1 SIFT特征 ８
2.3.1.2 AKAZE特征 １１
2.3.2 特征匹配算法 １５
2.3.2.1 暴力匹配 １５
2.3.2.2 近似最近邻匹配 １６
2.3.2.3 级联哈希匹配 １８
2.3.2.4 快速哈希匹配 ２０
2.4. 误配点过滤算法 ２３
2.5. 运动恢复结构 ２６
2.5.1 增量式SFM ２７
2.5.2 全局式SFM ３０
2.5.3 光束平差法 ３４
3 稠密三维点云的生成 ３７
3.1 MVS基本概念 ３８
3.2 密集点云生成原理 ３８
4 网格重建细化处理 ４２
4.1 三角网格 ４３
4.2 网格重建 ４３
4.2.1 网格重建基础知识 ４３
4.2.2 四面体网格的重建 ４５
4.3 网格细化算法 ５０
4.4 网格纹理 ５４
4.4.1 纹理映射 ５４
4.4.2 基于面的纹理映射 ５５
4.4.3 基于缝隙的纹理优化 ５５
5 基于图像序列的植物三维重建系统 ５９
5.1 实验环境及操作系统 ５９
5.2 系统结构及功能 ５９
5 *51今日免费论文网|www.jxszl.com +Q: ￥351916072￥ 
.3 实验结果 ６１
6 总结与展望 ６４
6.1 本文工作总结 ６４
6.2 未来工作总结 ６５
致谢 ６６
参考文献 ６６
基于图像序列的植物三维重建技术
摘 要
三维重建一直是计算机视觉领域的热门话题，在交通、医疗、虚拟现实中都已经有了广泛应用，而在农业方面仍有欠缺。为了弥补相关方面的研究，本文将采取几种不同的三维重建处理方法互相进行对比，从而获取到相对较优的方案，在基于单目视觉的基础上来实现三维点云的重构。首先需要获取到目标植株的360°全方位的序列图像，此后通过运动恢复结构的重建原理（structure from motion，SFM）得到期望物体的稀疏点云，由于稀疏点云中所包涵的三维空间数据较少，仍必需通过多视点立体视觉的方法（multiple view stereo，MVS）来获取包含了较多三维数据信息的稠密点云，利用得到的稠密点云进行三角网格化，最后获取可视化性较强的原三维空间影像，并通过可视化较强的交互界面来进行每个步骤的展示。整个过程中，本文通过采用openMVG的理论架构和算法思想，来实现稀疏点云的生成，并在其关键步骤中使用多种当前较为流行的算法来进行实现；之后再对获取到的稀疏点云数据，利用openMVS的理论架构和算法思想，来实现稠密点云的生成以及网格纹理化处理，最终得到目标的三维重建模型。最终，本文通过几种组合之间的比对，得到了以AKAZE特征和快速哈希匹配的提取算法，并结合基础矩阵与RANSCA过滤的估计算法，以增量式SFM算法结合基于深度信息的融合方式，得到最优的三维重建效果的结论。本文的实现平台是建立在Qt、VS2017及opencv之上完成的，其中还有引用到Ceres、CGAL等的外源算法包。
关键字：三维重建；稠密点云；稀疏点云；openMVG；openMVS
3D RECONSTRUCTION OF PLANTS BASED ON IMAGE SEQUENCE
Student majoring in computer science and technology LI Sizhe
Tutor WU Yanlian
ABSTRACT
3D reconstruction has been a hot topic in the field of computer vision. It has been widely used in transportation, medical treatment and virtual reality, while there are still some deficiencies in agriculture.In order to make up for the related research, this paper will take several different 3D reconstruction processing methods to compare with each other, so as to obtain a relatively better scheme.The reconstruction of 3D point cloud is realized based on monocular vision.First need to get into the target plant 360 ° full sequence images, then through sports recovery principle of the reconstruction of the structure (the structure from motion, SFM) of the desired object sparse point cloud, due to the sparse point cloud contains 3D spatial data is less, still must through the multiview stereo vision method (multiple view stereo, MVS) to get the 3D data contains more information of the dense point cloud, dense point clouds obtained by using triangular mesh, and finally obtain the visual sex strong original threedimensional images,And through the visualization of a strong interactive interface to display each step.In the whole process, this paper adopts openMVGs theoretical framework and algorithm idea to realize the generation of sparse point cloud, and USES a variety of currently popular algorithms in its key steps.Then, for the sparse point cloud data acquired, the theoretical architecture and algorithm idea of openMVS are used to realize the generation of dense point cloud and the mesh texturing process, and finally the 3D reconstruction model of the target is obtained.Finally, through the comparison among several combinations, the extraction algorithm based on AKAZE features and fast hash matching was obtained in this paper, and the optimal 3D reconstruction effect was obtained by combining the estimation algorithm of basic matrix and RANSCA filtering, incremental SFM algorithm and the fusion method based on depth information.The implementation platform of this paper is based on Qt, VS2017 and Opencv, among which there are external algorithm packages referring to Ceres, CGAL and so on.

原文链接：http://www.jxszl.com/jsj/jsjkxyjs/605734.html