Construction Of Point Cloud Dataset Of Typical Mechanical Parts

The 3D point cloud dataset is the basis for the research on point cloud semantic segmentation based on deep learning,but there is currently a lack of datasets for semantic segmentation of point cloud data of mechanical parts.The technology is applied to the precondition of semantic segmentation of part point cloud data.Constructing a dataset for semantic segmentation of point cloud data is a very time-consuming task due to the need for point-by-point labeling and a series of processing tasks for point clouds.In order to explore the method of quickly constructing datasets that meet the needs of deep learning,this paper conducts analysis and research in the following five aspects.（1）The process of generating virtual data of part point cloud in cluttered scene is designed.In order to obtain high-quality datasets that are close to the real environment,a point cloud data generation method based on Open Scene Graph,Bullet and osg Bullet is proposed.Specifically,Solidworks is used to create 3D models of common parts according to the real part size;a virtual scene containing a variety of part models is created using the 3D rendering engine Open Scene Graph.In order to simulate the stacking of parts,the physics engine Bullet is used to give the models gravity and collision detection and other mechanics.Features;randomly generate scattered scenes through parts collision,and obtain virtual point cloud data through uniform sampling.（2）The related algorithms of point cloud preprocessing are studied.The point cloud data collected in the real environment contains a lot of noise points,which will affect the subsequent point cloud processing.In order to effectively remove the noise in the point cloud,an improved point cloud denoising algorithm based on statistics is proposed,which organizes the point cloud data through the Kd-tree structure,and then effectively removes the noise points according to the distribution law of the average distance of the neighborhood of the point cloud data.General point cloud data contains a large number of points.In order to take into account the training efficiency and segmentation accuracy of deep learning,it is necessary to reduce the size of the point cloud on the basis of retaining the original structure of the point cloud.The downsampling method can specify the size of the point cloud on the basis of preserving the original point cloud structure and edge to the greatest extent.（3）The method of rapidly constructing point cloud dataset is studied.In order to improve the efficiency of point cloud data annotation,a traditional point cloud segmentation algorithm is proposed to replace the operation of manual point selection.First,the random sampling consistency algorithm is used to remove the part of the workbench where the parts are placed;then the LCCP segmentation algorithm is used to achieve the segmentation between parts;data enhancement methods such as changing the density of point cloud data,rotation and translation,and adding noise are proposed;After standardization,a point cloud dataset is generated according to the format of the S3 DIS standard dataset.（4）Based on the rapid construction process of point cloud datasets in cluttered scenes,a visual point cloud labeling software has been developed.This software can generate virtual point cloud data of cluttered scenes,and quickly construct the original point cloud data to meet the requirements of deep learning.At the same time,manual adjustment of the segmentation results of point cloud data can be realized.（5）The network structure of Point Net and Point Net++ is studied.The difficulties brought by the characteristics of point cloud data to deep learning technology in processing point cloud data are analyzed.The structural part of the Point Net network that can solve the difficult point cloud data processing is studied,and the structural design of the improved version of the Point Net++ network,such as encoding and decoding,is studied in detail.The random sampling algorithm is used to improve the training efficiency of Point Net++,and the loss function is redesigned..Finally,Point Net++ is used to verify the rationality of the rapid construction process of the point cloud dataset.