基于混合注意力机制和相关性估计网络的点云配准算法

A Point Cloud Registration Algorithm Using a Mixed Attention Mechanism and Correlation Estimation Network

点云配准是对同一物体上采集到的点云数据进行精确匹配,然而传统方法计算成本高,配准精度差.基于神经网络的算法也存在噪声干扰,在类别未见的点云数据上应用时效果不佳.为解决这一问题,本文提出了一种基于混合注意力机制和相关性估计网络的点云配准算法.考虑到点云内部特征的复杂性和点云对变换的随机性,提出一种混合注意力机制来提取关键特征信息,利用残差的方式进行连接,可以得到更具鲁棒性的点云特征.通过相关性估计网络对点云特征进行非线性激励,可以提高表达能力,获取点云对之间更紧密的相关性.在人工合成数据集ModelNet40和真实数据集ICL-NUIM上的仿真实验结果表明,本文算法在大尺度仿射变换下,对掺杂噪声、类别未见点云数据的配准精度有显著的提升,证明了其有效性.

Point cloud registration aims to accurately match point cloud data collected from the same object.However,traditional methods encounter high calculation cost and poor registration accuracy difficulties.Algorithms based on neural networks have not been effective in handling noisy and unseen point cloud data.To address this problem,a point cloud registration algorithm using a mixed attention mechanism(MA)and correlation estimation(CE)network is proposed.Considering that point clouds have complex internal features and random transformations,the MA is designed to fuse and extract the key features,which can be made more robust through a residual connection.Subsequently,the point cloud features are better expressed with the nonlinear excitation of the proposed CE network,which can obtain a closer correlation between point cloud pairs.Experimental results for the artificially synthesized dataset ModelNet40 and real dataset ICL-NUIM show a significant improvement in the registration accuracy of noisy and unseen point clouds under a large affine transformation,which demonstrates the effectiveness of the proposed algorithm.