Research On Point Cloud Completion Algorithm Based On Deep Learning

Point clouds are widely used in 3D data processing because of their ability to render objects in a fine detail.In recent years,with the advancement of 3D scanning and sensor technology,acquiring point cloud data has become easier and easier.However,during the actual acquisition of point cloud data,the acquired point clouds are often sparse and incomplete due to the target object being partially obscured,in relative motion or scanning equipment,and the geometry obtained directly or resampled also differs greatly from the real object.As an important research direction in the field of computer vision,point cloud complementation aims to recover the missing geometric and semantic information from incomplete point cloud data,and perform shape completion on the fragmented point clouds,so as not to affect various downstream tasks such as object recognition and detection,surface reconstruction,and robot operation.Therefore,the demand for point cloud completion is increasing,and it shows increasing application value and prospect in the fields of autonomous driving and medical image processing.In this paper,we propose a novel deep learning-based point cloud completion method,which takes the classical point cloud completion algorithm PCN as the base model,tries to introduce WGAN and Transformer technology,and innovatively designs an integrated encoder Encoder to combine the multi-layer perceptron and Transformer.The WGAN correction module based on the attention mechanism can better achieve the point cloud completion effect in the extreme case of sparse input point clouds or very small number of points;the decoder introduces PointNet++ feature extraction layer and multi-head external attention mechanism to better utilize the global features of point clouds obtained by the above modules for featurebased complete dense point cloud generation network,so that the output point cloud is closer to the real point cloud.To verify the effectiveness of the method,experiments are conducted on the ShapeNet dataset and significant performance improvements are achieved.In the commonly used judging criteria of point cloud completion effect,Chamfer Distance(CD)and Earth Mover Distance(EMD),when the number of points of input stump point cloud is 2048,the experimental results show that CD and EMD are 9.72 and 0.58,respectively,which are 7.95%and 1.69% lower compared with PCN;when the number of points of input partial point cloud is 2048,the experimental results show that CD and EMD are 9.72 and 0.58,respectively,which are 7.95% and When the number of input points of the partial point cloud is 16,the experimental results show that the CD and EMD are 15.17 and 0.89,which are 35.71% and 27.59% lower than the PCN,respectively.The uniformity and structural recovery of the completion results are better than the existing point cloud completion methods,which are closer to the real object point clouds.The 3D point cloud complementation algorithm proposed in this paper contains key technologies such as point cloud feature extraction,convolutional neural network and visual attention mechanism.