融合特征交互和点匹配增强的无监督点云配准算法

打开文本图片集

关键词：点云配准；特征交互；交叉注意力；点匹配增强；无监督中图分类号：TP391.4 文献标识码：Adoi：10.37188/OPE.20253320.3281 CSTR：32169.14.OPE.20253320.3281

Abstract： To address the issue of partial point mismatches in point cloud registration caused by outliers， partial overlap，and geometrically similar but non-corresponding points，this paper proposed an Unsupervised Point Cloud Registration Algorithm Integrating Feature Interaction and Point Matching Enhancement. First，a feature fusion module was developed to perform interactive integration between features of the source and target point clouds，and to fuse the resulting features with those extracted at the corresponding positions in the previous layer，thereby enhancing feature representation capability. Second，a graph Convolutional Network-Transformer fusion module was designed，in which graph convolution was employed to extract local geometric information，while the self-atention mechanism of the Transformer was used to capture global contextual information. A cross-attention mechanism was further incorporated to achieve effective feature interaction between point clouds.Finally，a point matching enhancement module was introduced，which established point correspondences by jointly considering the Euclidean distance of point features and the similarity of their local neighborhoods. The proposed algorithm was evaluated on the ModelNet4O（with noise），7Scenes，ICL-NUIM，KITTI，and ScanObjectNN datasets. Experimental results demonstrate that，compared with the IFNet algorithm，the proposed method achieves reductions in root mean square error RMSE（R） of 31.93% ， 23.72% ， 19.76% ， 10.53% ，and 21.05% ，respectively，validating its superiority in both registration accuracy and robustness. Overall，the proposed algorithm exhibits excellnt performance in registration accuracy，generalization capability，and noise resistance， showing strong potential for real-world applications.

Key words： point cloud registration；feature interaction；cross-atention；point matching enhancement; unsupervised

1引言

随着基于深度学习的点云配准技术的快速发展，点云配准在自动驾驶[2、医学影像处理[3]机器人导航[4]、以及地图构建[5]等领域得到了广泛应用。（剩余26698字）