Research On Point Cloud Completion Algorithm Based On Deep Network

In practical application scenarios such as 3D perception and robot navigation,the representation of data gradually develops from 2D images to 3D models.Among them,point cloud is favored by people as a form of expression with high flexibility and strong information effectiveness.However,due to the limitation of viewing angle and occlusion,the3 D data collected in real scenarios is usually sparse and incomplete,which leads to the loss of geometric and semantic information.The main challenge of point cloud application is how to recover the original shape information from a partial one.With the rise of artificial intelligence technology,point cloud completion methods based on deep learning gradually replace traditional methods and become the mainstream research direction.The current deep completion network adopts the ”encoder-decoder” structure as the main architecture,but the disorder of the point cloud data makes it difficult for the encoder to effectively obtain the topological details and the geometric structure of discrete points,which reduces the quality of the reconstructed point cloud.To tackle this problem,this paper uses the approximation method of high-dimensional feature similarity to propose a point cloud shape completion network based on twin encoders and then proposes a novel deep network named as the fine shape of point cloud completion network based on feature approximation.The main research contents of this paper are as follows:(1)Since it is difficult for the encoder to effectively obtain the topological details and the geometric structure of the discrete points,this paper proposes a point cloud shape completion network based on the siamese encoder.Therefore,this paper uses the embedding feature of the pre-trained complete point cloud auto-encoder network as the true value of the feature codeword to train the siamese network,so that the other twin encoder can produce similar output when the input is the defective point cloud.Through the training stage,the coding feature approximates the encoding feature of the complete point cloud.In this way,the encoder network can capture compact shape information more beneficially,which will be used by the decoder to reconstruct the point cloud later.Then through the decoding of the pre-trained auto-encoder to process the narrowed feature code words to achieve the task of point cloud completion.(2)The coding process of the network compresses the point cloud data,which the detailed structural information of the incomplete input point cloud is inevitably lost.We propose an innovative fine shape of the point cloud completion network based on feature approximation to address this issue.This network makes full use of prior information such as the incomplete input point cloud,the output point cloud of the intermediate network,and the latent embeddings after feature approximation,and then designs a refinement unit to restore precise point cloud shape by integrating this prior information effectively.Such a network structure can not only adjust the position of the point in the prior point cloud information but also continuously up-sample the output point cloud by using the iterative recursive network structure to obtain the completion result of the required resolution.Qualitative and quantitative experiments on open datasets demonstrate that point cloud shape completion network based on siamese encoders can recover the general shape of the partial input,and fine shape of point cloud completion network based on feature approximation can effectively reconstruct the detailed information of the object,which demonstrate that our proposed network outperforms the state-of-the-art point cloud completion methods.