随着人类社会生产发展，对信息需求的类型逐渐丰富，三维数据的研究受到众多学者的关注。三维数据包含极其丰富的信息，蕴含人类对世界的认知，利用计算机技术可以还原实体数字化的三维数据表示。三维重建技术是文物保护领域实现文物数字化保护和传承的核心技术之一，点云配准技术则是三维重建流程中的关键步骤。传统的点云配准方法对待配准的点云初始位置要求较高，计算效率低下，造成点云配准的实时性差，且配准精度有待提高。针对这些问题，本文设计了一种基于深度学习的点云配准算法，为验证算法的有效性，使用不同样本的点云数据进行了实验对比。具体研究内容及创新成果如下：
（1）本文针对点云特征提取，设计了一种点云图卷积网络，简称PGCN网络。PGCN网络通过构建点云的局部邻域图结构实现了对点云局部特征的提取，优化PointNet网络的点云全局特征的提取方法，剔除了原有网络的T-Net模块，提高网络计算的效率。
（2）本文针对点云配准，将经典的Lucas-Kanade光流算法改造为适用于神经网络训练的光流算法，使用反向组合的方式，提高配准的计算效率。最后将PGCN点云特征提取网络视作成像函数，将改造后的Lucas-Kanade光流算法集成到PGCN网络中，构成PGCNLK点云配准网络，实现点云配准精确度提升。
（3）在实验分析方面，使用训练集同分布样本的点云数据和文物样本点云数据进行实验，验证本文设计的PGCNLK点云配准网络的有效性，结果表明，PGCNLK网络实现了点云的精确配准。同时与ICP算法进行了对比，证明了PGCNLK网络在配准时间、配准精度等指标上的优势。

With the development of human society and innovative technology, the research of three-dimensional data has attracted the attention of many scholars. Three-dimensional data contains extremely rich information, which is in line with human cognition of the world. The use of computer technology to restore the world represented by three-dimensional data can realize the digitization of the industry. In the field of cultural relics protection, digital technology is beneficial to the protection and inheritance of cultural relics. 3D reconstruction technology is one of the core technologies to realize the digitalization of cultural relics. Among them, point cloud registration technology is a key step in the 3D reconstruction process.
This thesis analyzes the traditional point cloud registration algorithm and gets the following conclusions: First, the traditional point cloud registration method requires strict initial position of the point cloud to be registered. Second, the calculation efficiency is low, which affects the real-time performance of the point cloud registration. With consideration of the above problems, it is feasible to improve accuracy and efficiency of registration algorithm. This thesis proposes a point cloud registration algorithm based on deep learning algorithm. In order to verify the effectiveness of the algorithm, point cloud data from different data sets are used for experimental comparison. The specific research content and results are as follows: 
(1) In terms of point cloud feature extraction, this thesis proposes a point cloud graph convolutional network, referred to as PGCN network. The PGCN network realizes the extraction of the local features of the point cloud by constructing the local neighborhood graph structure of the point cloud, and realizes the extraction of the global feature of the point cloud by referring to the core idea of the PointNet network design. When extracting the global feature, the T- of the PointNet network is eliminated. Net module enhances the efficiency of network calculation.
(2) In terms of point cloud registration, this thesis transforms the classic Lucas-Kanade optical flow algorithm into the Lucas-Kanade optical flow algorithm suitable for neural network training, and uses the reverse combination method to improve the calculation efficiency of the registration. Finally, the PGCN point cloud feature extraction network is regarded as an imaging function, and the reconstructed Lucas-Kanade optical flow algorithm is integrated into the PGCN network to form a PGCNLK point cloud registration network to realize the point cloud registration task. 
(3) In terms of experimental analysis, using the same point cloud data of the training set and the point cloud data of cultural relic samples to conduct experiments to verify the effectiveness of the PGCNLK point cloud registration network designed in this thesis. The results show that the PGCNLK network realizes the point cloud. The precise registration. At the same time, it was compared with the ICP algorithm, which proved the advantages of the PGCNLK network in terms of registration time and registration accuracy.