Research On 3D Point Cloud Completion And Registration Based On Deep Learning

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 lidar,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 have 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 completion and registration problems in 3D point clouds.The content is as follows:(1)Existing point cloud completion methods focus on reconstructing the overall structure,but the protruding points or small irregular surfaces are difficult to predict,and the integrity of the point cloud is often affected by discontinuous material surfaces or rough surfaces.Due to the influence of sensor resolution,it is difficult for existing methods to achieve good completion accuracy.To address this problem,we propose a new end-to-end neural network for point cloud completion.To avoid uneven point cloud density,regular voxel centers are chosen as reference points.Encoder and decoder are designed with Patchify,Transformer and Multilayer Perceptron.The implicit classifier is incorporated into the decoder to mark valid voxels,which can be diffused in the completion,using a newly designed loss function,training the classifier to learn contours,which helps to identify difficult-to-judge points for diffusion.The effectiveness of the proposed model is verified in comparative experiments with the state-of-the-art on the indoor Shape Net dataset and the outdoor KITTI dataset,which shows that this method can more accurately predict points with rich details and uniform point cloud distribution point cloud coordinates.(2)Existing point cloud registration methods have a large performance loss in low-overlap scenarios,and most methods have poor generalization performance.In this paper,we design a new registration model that performs better in low-overlap scenes and possesses strong generalization performance.On the one hand,to address the registration problem in low-overlap scenarios,we propose a novel fully convolutional network to search for superpoints in overlapping regions and generate feature descriptors at the superpoints.The goal of this network is to filter points in nonoverlapping regions as well as smooth regions.On the other hand,we introduce a rotation-invariant convolution strategy for the full convolutional model,which makes the features extracted by the network invariant to rotation,thereby improving the generalization of features.We test our method on 3DMatch,3DLo Match,KITTI and ETH datasets and compare with SOTA methods.Experimental results show that the method can achieve the best performance on the low-overlap registration task,and it performs well in different scenes detected by different sensors.