Cy-CNN: cylinder convolution based rotation-invariant neural network for point cloud registration

作     者：Hengwang ZHAO Zhidong LIANG Yuesheng HE Chunxiang WANG Ming YANG 

作者机构：Department of Automation Shanghai Jiao Tong University Research Institute of Robotics Shanghai Jiao Tong University 

出 版 物：《Science China(Information Sciences)》 (中国科学:信息科学(英文版))

年 卷 期：2023年第66卷第5期

页      面：73-87页

核心收录：

学科分类：12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 081104[工学-模式识别与智能系统] 08[工学] 080203[工学-机械设计及理论] 0835[工学-软件工程] 0802[工学-机械工程] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：supported by National Natural Science Foundation of China (Grant Nos. 62173228  61873165) 

主　　题：deep learning point cloud registration rotation invariant computer vision 

摘      要：Point cloud registration is a challenging problem in the condition of large initial misalignments and noises. A major problem encountered in the registration algorithms is the definition of correspondence between two point clouds. Point clouds contain rich geometric information and the same geometric structure implies the same feature even if they are in diferent poses, which motivates us to seek a rotation-invariant feature representation for calculating the correspondence. This work proposes a rotation-invariant neural network for point cloud registration. To acquire rotation-invariant features, we firstly propose a rotationinvariant point cloud representation(RIPR) at the input level. Instead of using the original coordinates,we propose to use point pair features(PPF) and the transformed coordinates in the local reference frame(LRF) to represent a point. Then, we design a new convolution operator named Cylinder-Conv which utilizes the symmetry of cylinder-shaped voxels and the hierarchical geometry information of the surface of3D shapes. By specifying the cylinder-shaped structures and directions, Cylinder-Conv can better capture the local neighborhood geometry of each point and maintain rotation-invariance. Finally, we combine RIPR and Cylinder-Conv to extract normalized rotation-invariant features to generate the correspondence and perform a diferentiable singular value decomposition(SVD) step to estimate the rigid transformation. The proposed network presents state-of-the-art performance on point cloud registration. Experiments show that our method is robust to initial misalignments and noises.