基于偏置注意力边特征的鲁棒点云配准算法

Robust Point Cloud Registration Algorithm Based on Biased Attention Edge Features

摘要:

刚性点云配准已成为机器人和计算机视觉领域的一项基本任务。为了解决基于空间距离的迭代最近点（ICP）对初始姿态敏感及噪声不鲁棒的问题，提出了一种基于偏置注意力边特征的鲁棒点云刚性配准算法。该方法首先使用动态图神经网络学习混合特征，融合空间坐标、局部几何信息和全局特征。然后利用偏置注意力模块增强点云特征的显式上下文信息，有效减少了点云变换和噪声带来的影响。其次，通过定义在混合特征上的退火参数和行列交替归一化得到点云之间对应关系的软匹配矩阵。最后采用奇异值分解估计变换，实现了端到端的点云配准算法，提高了配准的精度和鲁棒性。在ModelNet40数据集上进行的大量实验表明，该方法优于传统方法和基于学习的方法，取得了较好的性能。

Abstract:

Rigid point cloud registration is a fundamental task in robotics and computer vision. To address the sensitivity of traditional spatial distance-based iterative closest point （ICP） algorithms to initial transformations and noise， this paper proposed a robust rigid registration algorithm based on bias-attention edge features. First， a dynamic graph convolutional neural network was used to extract hybrid features that fuse spatial coordinates， local geometry information and global feature. Next， a biased attention module was utilized to enhance the explicit point cloud context information， which effectively reduced the effects of point cloud noise. Then， the algorithm computed soft matching matrix for the corresponding relationship between the point clouds by defining annealing parameters on mixed features and performing alternate normalization of rows and columns. Finally， the transformation between point cloud is estimated using singular value decomposition in an end-to-end framework， enhancing both the accuracy and robustness of registration. Extensive experiments conducted on the ModelNet40 dataset show that the proposed algorithm out perform both traditional and state-of-the-art learning-based methods.