一种使用伪对应点生成的3D点云配准方法

Deep Artificial Correspondence Generation for 3D Point Cloud Registration

针对三维重建过程中点云配准存在的挑战性问题(如寻找对应点困难等)展开研究,充分利用源点云和目标点云的几何信息,提出了一种基于交叉注意力和伪对应点生成机制的点云配准方法——深度伪对应点生成(DeepACG)。该方法采用三级网络模型,第一级是深度特征编码模块,利用交叉注意力机制交换和增强两片待配准点云之间的上下文和结构信息;第二级是伪对应点生成模块,基于软映射关系加权合成伪对应点;第三级为对应点加权和离群点过滤模块,赋予每个对应点对不同的权重值并剔除概率较低的离群点。在合成和真实数据集上进行大量实验,DeepACG方法在室内真实数据集3DMatch上的配准召回率达到92.61%;在数据集ModelNet40上进行目标未知的局部点云配准实验,旋转矩阵和平移向量的均方根误差分别降至0.016和0.00009。实验结果表明,DeepACG配准精度高,鲁棒性强,配准误差低于当前主流的配准方法。

To address the challenging problems of point cloud registration in 3D reconstruction(e.g.,difficulty in finding corresponding points,etc.),this paper proposes a point cloud registration method based on cross-attention and artificial correspondence generation mechanism,Deep Artificial Correspondence Generation(DeepACG),by fully utilizing the geometric information of the source and target point clouds.Our method adopts a three-stage network model.The first stage is the deep feature encoding module,which exchanges and enhances the contextual and structural information between two unaligned point clouds using the cross-attention mechanism.The second stage is the artificial correspondence generation module,which synthesizes the artificial correspondences by weighting the soft mapping.The third one is the correspondence weighting and outlier filtering module,which assigns different weights to the correspondence pairs and rejects them with a small probability.Extensive experiments are conducted on both synthetic and real-world datasets.Our method achieves a registration recall of 92.61%on the real-world indoor dataset 3DMatch,and we execute unseen partial registration experiments on ModelNet40,reducing the root mean square error of the rotation matrix and translation vector to 0.016 and 0.00009,respectively.Experimental results show that DeepACG has higher registration accuracy and robustness,and its alignment error is lower than that of the existing mainstream registration approaches.