Dense Point Cloud Completion Based On Deep Learning

Point cloud completion aims to reconstruct complete point clouds from partial point clouds,which is widely used in various fields such as autonomous driving and robotics.Most of the existing methods are sparse point cloud completion,where the number of point clouds after completion is relatively small and the details are insufficient.PCN model is the first to propose dense point cloud completion and the number of complete point clouds is about 8 times that of sparse point clouds.The dense point clouds after completion have more detailed features,which provides the core basis for the follow-up point cloud processing tasks.However,this kind of model mainly focuses on the global feature information of point clouds and pays less attention to the dependency between point clouds and their local feature information.To solve the above challenges,this paper explores the field of point cloud completion and puts forward the following methods:First,this paper proposes a novel dense point cloud completion network called NDPC,which integrates Self-attention mechanism with the fusion of global and local feature information of point clouds to achieve shape completion.First of all,we apply the Self-attention mechanism to point clouds’ global feature extraction to make it emphasize the dependency between different points.Secondly,we adopt the multi-stage completion where we obtain coarse point clouds at the first stage and combine local and global feature information to achieve the completion of dense point clouds.The quantitative and qualitative evaluations of experiments demonstrate that N-DPC model has achieved better performance on the ShapeNet dataset compared with existing state-of-the-art methods.It leads the PCN model by 5%in the Earth Mover ’s Distance(EMD)and shows good robustness for different missing ratios of point clouds.What’s more,experiments on the KITTI car dataset show that N-DPC model is also valid for point cloud completion in real point cloud scenarios.Besides,as a dense point cloud completion algorithm,N-DPC also leads the PCN model by 7%in the Chamfer Distance(CD)of the Completion3D sparse point cloud dataset.Second,this paper proposes N-DPC-GAN based on the N-DPC model and generative adversarial network.The model aims to use the generative adversarial network to construct a one-to-one mapping relationship between the high-dimensional global feature vector of partial point clouds and the high-dimensional global feature vector of target point clouds(Ground Truth)to further "optimize" the high-dimensional global feature vector of partial point clouds.Experiments show that N-DPC-GAN model achieves comparable results on the ShapeNet dataset.The model has significant advantages when the lack of a large proportion of point clouds results in a large loss of spatial structure,and the greater the loss percentage,the more obvious the utility.When the percentage of missing point clouds is 80%,the performance of N-DPC-GAN model is 10%better than that of the N-DPC model.Finally,we further explore the application of dense point cloud completion in point cloud classification.The results show that N-DPC-GAN model outperforms other methods in the case of a large missing proportion of point clouds and the classification accuracy of complete point clouds is as high as 86.5%under the condition of 80%missing point clouds.