Research On The Algorithm Of 3D Point Cloud Shape Repair Based On Generated Adversarial Network

Nowadays,3D reconstruction technology has been greatly developed.Whether it is the improvement of hardware or the progress of reconstruction algorithms,the generated point cloud model is both dense and accurate,but there are still some inevitable problems.In the real scanning environment,due to line of sight blocks or improper operation of technicians,there will be data loss in the point cloud actually obtained.The incomplete point cloud will have a serious impact in the reconstruction process,resulting in the inability to effectively obtain the complete structure and surface morphology of the model,which will cause inconvenience to subsequent applications.Therefore,the repair of missing data has become the key to the research and application of 3D models.The traditional methods based on curvature features,mutual constraints between grids or neighborhood information of hole area can not meet the needs of large area data missing repair;the methods based on database prior knowledge or predefined rules rely too much on manual design,which is inefficient.With the development of deep learning and 3D database,the algorithm of 3D feature extraction based on deep neural network has made a breakthrough in 3D field.Therefore,in view of the data missing phenomenon of point cloud model,we have studied the repair of large-area data loss of the point cloud based on the generated adversarial network.The main work includes the following three points:(1)In order to solve the problem of missing point cloud data,we introduces the encoder-decoder network into the field of 3D point cloud shape repair,through deep learning powerful learning ability,using the whole information of point cloud model to effectively repair the missing area structure.Compared with the 3D-EPN algorithm based on voxel CNN,the algorithm can effectively repair the point cloud model,but the repair results are uneven distribution.(2)In order to solve the problem of uneven distribution of the point cloud repaired by the AutoEncoder network,we uses the idea of generative adversarial networks for reference,improves the AutoEncoder network,and proposes a point cloud shape repair algorithm based on the generative adversarial network.By comparison,the algorithm further improves the ability of repairing point cloud shape of AutoEncoder network,and makes the distribution of repaired point cloud more uniform.Compared with other point cloud shape repair algorithms based on deep learning,the proposed algorithm achieves the same or even slightly better performance on the test data.(3)The RTAB-Map algorithm is used to collect point cloud data through Kinect sensor in the simulation environment and the real scene.Then the PCL point cloud database algorithm is used to segment the scene point cloud,remove outliers,and take down samples to obtain the three-dimensional model of data loss.Finally,the point cloud of data loss will be generated adversarial network shape repair.