Research On Completion Algorithm Of 3D Point Cloud Data

With the development of robotics,autopilot,virtual reality and other fields,the requirements for the ability of machines to perceive the surrounding environment are becoming higher and higher.As an important sensor that can sense the depth information of the surrounding environment,the processing of the data format directly obtained by the lidar,that is,the point cloud data,is also increasingly important.In reality,the data collected by the laser radar is usually not a complete point cloud model.Due to occlusion and interference,the point cloud data scanned by the lidar will be partly missing,resulting in the loss of geometric information of the object.The incompleteness of the point cloud data affects the further processing of the point cloud.Therefore,completing partial point clouds into complete point clouds has important value for downstream applications such as classification,segmentation,target detection,etc.This thesis discusses some milestone networks in the field of point cloud processing.These networks directly process point cloud data instead of converting it to data in other formats before processing.PointNet is the first network to directly process point cloud data,which solves the problems caused by the rotation invariance,unstructured,and disorder of point clouds,but it ignores local features in the process of processing.Point Transformer migrates the Transformer in the field of natural language processing to the field of point cloud,and solves the problem that PointNet loses local features.FoldingNet has designed an auto-encoder,which proves the feasibility of restoring point clouds from high-dimensional features,and provides a reference for the structure of point cloud completion network.And PCN completes the task of point cloud completion.Based on the network structure of FoldingNet,this thesis proposes an end-to-end point cloud completion network.The network adopts an encoder-decoder structure.The encoder is used to extract the high-dimensional features of the input incomplete point cloud,and the decoder is used to reconstruct the missing part of the point cloud according to the high-dimensional features.The main innovations and improvements are as follows: Firstly,most point cloud completion networks directly output complete point clouds.The network proposed in this article only predicts missing point clouds,and then splices the output and input to obtain complete point clouds.This can reduce the difference between the completion results and the Ground Truth,and reduce the number of network parameters.Secondly,the network uses Point Transformer as the encoder,rather than the structure similar to multi-layer perceptron adopted by FoldingNet.In this way,the global and local features of the point cloud can be extracted together,which can improve the feature extraction ability of the encoder.Finally,I replaced the two-dimensional lattice in the FoldingNet decoder with the threedimensional lattice.According to the high-dimensional features,the three-dimensional lattice is deformed twice to make it close to the shape of the missing part.Finally,I replaced the two-dimensional lattice in the FoldingNet decoder with the threedimensional lattice.According to the high-dimensional features,a three-dimensional lattice is deformed twice to make it close to the shape of the missing part.This can avoid the conversion from 2D to 3D and improve the effect of outputting point clouds.This thesis uses ShapeNet as the dataset,uses kNN algorithm to simulate occlusion,and mainly uses the point cloud data of the car to do the experiment.For the missing part of the point cloud,the CD loss of the network is 4.104 × 10-2,about 8% lower than FoldingNet.I also used point clouds with different densities and degrees of occlusion,as well as other categories of point cloud as datasets for experiments,which confirmed the availability of the proposed network in the above data.