Strong Generalization Fully Convolutional Network For Low-overlap Point Cloud Registration

With the rapid development of 3D information collection technology,3D sensors are increasingly being used in scientific research and daily life.Various 3D sensors,such as Li DAR,RGB-D cameras,and 3D scanners,can provide rich geometric and proportional information for the three-dimensional data obtained,which provides the possibility for a more comprehensive understanding of the surrounding information of objects.3D data has a wide range of applications in areas such as autonomous driving,robotics,and medical care.Point cloud is the most common data form in 3D data.It is a collection of points used to express the spatial distribution and surface characteristics of objects in the same spatial reference system.Many scholars at home and abroad have proposed many 3D point cloud technologies for classification,segmentation,completion,configuration and other research.Based on the deep learning theory,this paper studies the registration problems in 3D point clouds.The research background of 3D point cloud registration is that given a pair of point clouds detected by sensors in different poses,the rotation parameters and translation parameters are solved so that one of the point clouds is aligned with the other point cloud after rotation and translation.Existing point cloud registration models suffer from large performance loss in low-overlap scenarios,while the generalization ability of most models are weak.In this article,we design a new model for point cloud registration pursing better low-overlap performance and generalization ability.On the one hand,to solve the registration problem in low-overlap scenes,we propose a novel full convolutional network searching for super points located in the overlapping region and generating feature descriptors at the super points simultaneously.The new network aims at extracting points beyond nonoverlapping or smooth regions.On the other hand,we introduce a rotation-invariant convolution strategy for the fully convolutional model so that the extracted feature descriptors have rotation invariance,which improves the generalization performance of the features.Our method is tested on 3DMatch,3DLo Match,KITTI,and ETH,and compared with the state-of-the-art methods.The experimental results demonstrate that our method can achieve the best performance in low-overlap registration tasks,and it performs well across unseen scenarios with different sensor modalities.