Research On Segmentation And Geometric Element Feature Recognition For Point Cloud

In modern industrial manufacturing,high-quality machining of workpieces is crucial for companies.For the production of some high-precision components in industries such as aerospace,automotive,and machinery,analyzing and detecting the shape,size,and geometric features of workpieces is very important.These geometric feature parameters can have a significant impact on the functionality,performance,and quality of components.Traditional manual measurement methods often require a lot of time and labor costs,and are often unable to handle complex parts,making it difficult to meet the requirements of high efficiency and accuracy in large-scale production.With the rapid development of 3D scanning and point cloud segmentation technology,the cost of acquiring point cloud data is becoming lower,and the accuracy of point cloud data is becoming higher.Through point cloud segmentation technology,geometric primitives in the point cloud can be segmented,and relevant parameters of the primitives and workpieces can be calculated.Therefore,how to use point cloud data for primitive segmentation and geometric feature measurement has become a highly researched area.This article focuses on research on 3D point cloud segmentation and geometric primitive recognition,with a particular emphasis on addressing the difficult issues of confusing plane and cylindrical surface identification and unclear primitive boundary segmentation.The main work of the paper is as follows:(1)A fast detection method for surface primitives based on model fitting is proposed to address the problem of local areas of low curvature cylindrical surfaces being identified as planes on the part model.This method can accurately and quickly detect both planes and cylindrical surfaces.Firstly,the point cloud is divided into small granularity patches,and patch features are calculated to roughly identify plane patches or cylindrical patches.Then,neighboring plane patches are filtered based on filtering conditions of cylindrical patches,and patches with the same features are merged to obtain complete planes and cylindrical surfaces.The experimental results show that this method effectively solves the problem of local areas of low curvature cylindrical surfaces being mistakenly identified as planes.Compared with the popular Efficient Random Sample Consensus(eRANSAC)and connectivity-based cylinder detection(CbCD),this method shows good performance and does not have omissions or misidentification.At the same time,in the accurate segmentation of multiple connected cylindrical surfaces and the accuracy of surface parameters,the proposed method is superior to the other two.(2)A deep network based on the Transformer is proposed for instance segmentation of primitives in point clouds to address the problem of unclear boundary segmentation of primitives in neural networks.Unlike traditional convolutional neural networks,this paper uses Transformer modules to encode position information and local context information in the point cloud.Based on this method,it can better handle details and irregular shapes in point clouds.At the same time,in order to enhance the effect of boundary segmentation,this paper also trains a primitive discriminator to supervise whether two points come from the same instance.The discriminator can supervise the network output and provide more accurate boundary segmentation information.The experimental results show that compared with the current primitive instance segmentation network,the network proposed in this paper achieves excellent performance in primitive instance segmentation tasks,with improvements of more than 0.7 in two mIOU metrics and three AP metrics.Compared with the network without discriminator supervision,the proposed method has significantly improved performance in boundary segmentation and small object segmentation.(3)On the basis of the above work,a prototype system for point cloud element segmentation is designed,which realizes file reading,point cloud visualization,point cloud sampling,point cloud filtering,and two point cloud segmentation functions based on model fitting and Transformer deep neural network,which promotes the solution of confusion and unclear boundaries in point cloud element segmentation and recognition in actual industrial manufacturing,and also provides reference value for point cloud identification in other fields.