Research On Point Cloud 3D Reconstruction Technology Of Large Workpiece Based On Deep Learning

Large workpieces such as single blades of controllable pitch propellers,engine casings,and reduction gearboxes are the key components of modern industrial equipment.The three-dimensional reconstruction and measurement of them can provide a data basis for the optimization of the production process of large workpieces.With the advancement of photoelectric technology and computer image processing technology,point cloud 3D reconstruction technology has gradually become a research focus in the field of 3D reconstruction for large workpieces.However,the traditional point cloud 3D reconstruction method has low accuracy and efficiency in the reconstruction of the workpiece to be measured.To solve the above problems,this thesis proposes a point cloud 3D reconstruction system based on deep learning to improve the accuracy and efficiency of the point cloud 3D reconstruction.The main research contents of the thesis are as follows:(1)Aiming at the problem that the traditional point cloud 3D reconstruction method cannot guarantee the overlap rate between the point clouds to be registered,resulting in insufficient 3D reconstruction accuracy,a multi-point cloud 3D reconstruction method based on a global template is proposed.Convert the ideal CAD model format with the complete structure of the workpiece to be tested into a template point cloud,introduce the template point cloud as a global template during the registration process of a single local point cloud,and gradually integrate multiple local point clouds through a similar patching method Register to the global template to obtain a fully registered model composed of local point clouds,and complete the 3D reconstruction of the workpiece to be measured based on the fully registered model.The complete 3D reconstruction experiment on the physical model shows that the method can improve the accuracy and efficiency of the point cloud 3D reconstruction.(2)Aiming at the large error and low efficiency of corresponding point search in existing point cloud registration algorithms,a template point cloud clipping network is proposed.Feature extraction and fusion are performed on the local point cloud and the template point cloud through the neural network structure,and the corresponding point set of the local point cloud is "cut out" from the template point cloud by using the fusion features of the two,and then the search for the corresponding points is realized.The corresponding point search method does not need to repeatedly calculate the local features of the point cloud,nor does it need to separately train specific local data.The corresponding point search experiment is carried out based on the point cloud data set,which shows that the network can significantly reduce the corresponding point search error and improve the corresponding point Search efficiency.(3)Aiming at the problems of low computational efficiency and narrow application range of the existing coarse point cloud registration algorithm,a rigid body transformation matrix parameter estimation network is proposed.The global features of the corresponding point set and the local point cloud are extracted through the neural network structure,and the global features of the corresponding point set and the local point cloud are the same after the complete registration,and the parameters of the rigid body transformation matrix in the registration are estimated.A point cloud coarse registration experiment based on a point cloud data set shows that the network can improve the efficiency of point cloud coarse registration,and at the same time,it can effectively calculate the rigid body transformation matrix between two point clouds with a large difference in the total amount of data points,Which expands the scope of application of deep learning algorithms in the field of coarse point cloud registration.