Algorithm And Research Of 3D Point Cloud Completion Based On Multiresolution Geometric Structure Information

With the widespread popularity of 3D scanner equipment and the rapid development of 3D vision technology,point cloud data has received special attention in academia and industry.Point cloud data is a set of vectors in a 3D coordinate system,which contains simple structure and rich information,including 3D coordinates,color,and direction.It is widely used in the fields of automatic driving,augmented reality,robots,such as detecting,tracking,and classifying objects.However,due to unavoidable factors such as light reflection,target occlusion,noise,the acquired 3D point cloud data may be missing or sparse.The point cloud completion has important research significance and application value for developing 3D vision technology.It is a key problem for high-resolution and efficient point cloud completion to use the geometric structure information of multi-resolution incomplete point clouds.Due to 3D point cloud has one more dimension than the image,multi-resolution 3D point cloud completion is facing more complex problems.Its research difficulties are mainly including:(1)Spatial distribution complexity.Point clouds are disordered,sparse,and irregularly distributed,which need to consider how to deal with these disordered data in computer.(2)Insufficient geometric information.Compared with the two-dimensional image,the point cloud lacks the texture in the two-dimensional image,resulting in the lack of geometric structure information,generating noise points and reducing the point cloud completion accuracy.(3)Existing discriminators cannot effectively distinguish between predicted point clouds and ground truth,which cannot improve optimization ability of discriminators and the accuracy of point cloud completion.In view of the above problems,this thesis has carried out research based on multi-resolution incomplete point cloud,and the main works are as follows:(1)To the research difficulties of spatial distribution complexity,a point cloud completion network based on graph structure is proposed.After constructing the undirected graph using the K-nearest neighbor strategy,the network further uses the multi-layer perceptron and the maximum pooling layer to solve the disorder of the point cloud.At the same time,it uses the skip connection in the feature decoding process to reduce the loss of feature information.The algorithm effectively extracts the geometric relationship between the neighborhood point and the center point,and improves the accuracy of the point cloud completion model.(2)To solve the problem of insufficient geometric information,this thesis designs a transformer-based network for point cloud completion.Firstly,a two-level biased attention mechanism is designed to capture the long-distance feature dependency and embed the spatial correlation between the features.Secondly,the two-level feature is concatenated to fully extract local information.Finally,the point pyramid decoder is used to generate a multi-resolution complete point cloud.The proposed method further proposes consistency loss to calculate the consistency of multi-resolution point clouds at the same resolution,to make the generated point clouds more uniform.(3)For inadequate discrimination ability.to further optimize the prediction results,this thesis proposes a projected generative adversarial network for point cloud completion,which projects the generated complete point cloud into the 2D camera coordinate system to determine its authenticity,and finally uses the projection discriminator to adversarial training generated model,to improve the accuracy of the 3D point cloud completion model.On the basis of the projection generative adversarial discriminator,this thesis also designs a skeleton feature enhancement module to infer the skeleton points of incomplete point clouds,captures its features as local geometric information,and embeds local information into the extracted features to generate fine complete point clouds.