本文为西班牙加泰罗尼亚政治大学（作者：Javier de la Rica Palma）的学士论文，共41页。

本项目源于学习神经网络工具的动机，以及在学位课程中所获得知识在工作中的应用，该工作将使用三维点云对神经网络进行训练。在这项工作中，我们创建了一个系统，该系统从以前的数据库中，利用从每个点云关键点获得的信息，计算出一个更大的数据集。本项目的目标是利用神经网路训练从两个三维点云中找出点之间的对应关系。一方面，如前所述，从不同的视点生成同一场景的三维点云数据库，然后用这些数据库比较所有点云的面片，以找出它们之间的对应关系。另一方面，生成一个神经网络，以预先计算的数据库作为输入进行训练，观察上述关键点的点云比较行为。总之，该项目的主要目标是研究神经网络在特定情况下的行为，为增强现实（AR）的应用提供可能。因此，本工作能够生成增强现实的应用，其中的三维对象还可以被插入到二维图像中。

The current project arises from themotivation of studying the tools of neural networks, along with the applicationof the knowledge acquired during the courses in the degree in a work thatbrings together the training of neural networks using threedimensional pointclouds. In the work, a system has been created, which from a previous database,computes a larger dataset making use of the information obtained from each ofthe point clouds key points. The objective of the project is to findcorrespondences between points from two three-dimensional point clouds usingneural networks training. On the one hand, as mentioned, the database ofthree-dimensional point clouds of the same scene has been generated fromdifferent points of view with ground truth, which later is used to comparepatches of all the point clouds, in order to find correspondences between them.On the other hand, a neural network is generated to be trained with thepreviously computed database as input, and observe the behavior of thecomparison of point clouds given the key points mentioned above. All in all,the main objective of the project is the study of the behavior of neuralnetworks in a specific case, for a possible application of augmented reality(AR). Therefore, the present work is a good basis for a possible alternativework, capable of generating an augmented reality application in whichthree-dimensional objects are inserted into two-dimensional images.

1.  引言

2. 理论基础

3. 研究现状

4. 研究方法

5. 研究结果

6. 预算

7. 结论

更多精彩文章请关注公众号：