Dynamic Graph CNN With Attention Module For Partially Overlapping Point Cloud Registration

3D point cloud registration is a basic task of pose estimation,3D reconstruction,SLAM and other computer tasks.For partially overlapping 3D point clouds with noise in different frames,it aims to find the best transformation that can align all point clouds in the same coordinate.In application,we allways get the partially overlapping point clouds for registration,due to the influence of multiple factors such as environment or dynamic object occlusion,equipment and environmental noise in the process of data collection.Because the non overlapped points and noise always invalidate the results of registration,so it is important to solve the impact of non overlapped points.Therefore,based on the partially overlapping point clouds registration failure,we proposed a deep learning point cloud registration method to solve it.The network divides point cloud registration into three modules.Firstly,deep graph point cloud feature extraction module is used to extract the topological structure features of point cloud;Secondly,an attention module is used to identify the overlapping point cloud;Finally,a pose optimization module based on maximum expectation and Gaussian mixture distribution is used to calculate the pose of registration results.In order to verify the effectiveness and robustness of our model,we trained with modelnet40.Experiments show that the network is effective,robust with noise,and most importantlly,improves the accuracy of registration tasks with partially overlapping point clouds.