基于双分支融合学习和二次图结构优化的点云配准算法

A point cloud registration algorithm based on dual branch fusion learning and quadratic graph structure optimization

目的:为了明确建模非结构化点云所蕴含的图结构以及缓解深度图卷积造成的匹配模糊。方法:首先,在局部邻域中构建富含中心点特征和邻居点特征信息的代理点。针对获得的代理点,采用自适应投票的方式学习代理点间的图结构。其次,将包含局部结构信息的节点特征与逐节点独立地学习到的特征融合,从而挖掘每个节点潜在的判别性特征。最后,通过最大化两个邻接矩阵之间的一致性来达到优化对应矩阵的目的,实现更精确的配准性能。结果:大量的对比实验结果显示,本文构建的模型具有最佳的配准效果,其中,在不可见类别的局部到局部的噪声点云配准的误差为RMSE(R)=0.4273,RMSE(t)=0.0035,CCD=0.0832。结论:基于双分支融合学习和二次图结构优化的匹配算法显著地提高了点云配准的性能。

Aims:This paper aims to paper aims to explicitly model the graph structure contained in unstructured point cloud and alleviate the matching confusion caused by depth graph convolution.Methods:Firstly,the proxy nodes rich in the features of center nodes and neighbor nodes were constructed in the local neighborhood.For the obtained proxy nodes,the graph structure between them was learned by adaptive voting.Secondly,the node features including local structural information were fused with the features learned independently node by node,thereby mining the potential discriminative features of each node.Finally,the objective of optimizing the correspondence matrix was achieved by maximizing the consistency between the two adjacency matrices for more accurate registration.Results:Extensive comparative experimental results revealed that the model constructed in this paper had excellent registration performance with the errors of partial-to-partial noisy point cloud alignment on unseen categories,RMSE(R)of 0.4273,RMSE(t)of 0.0035,CCD of 0.0832,respectively.Conclusions:The algorithm based on dual-branch fusion learning and quadratic graph structure optimization significantly improves the performance of point cloud registration.