3D Point Cloud Segmentation And Sphere Recognition Based On Deep Learning

In industry,the identification of spherical workpieces is easy to be affected by other impurities,resulting in the malpractice of wrong identification.By using point cloud identification based on deep learning to identify the feature of workpiece,the efficiency of industrial identification can be improved.There is an urgent need in industrial production.For Point Cloud identification,high-precision point cloud segmentation is the basis of point cloud identification.The disorder and non-structure of point cloud data make it difficult to segment point cloud.PointNet solves the disorder problem of point clouds.However,the frame pays more attention to the information of the point cloud than to the attribute relationship between the adjacent points of the point cloud,which leads to the loss of local features and the error of segmentation and recognition.At the same time,the point cloud of the sphere is easily affected by the noise and the object is incomplete.The feature of sphere is acquired by identification frame,which is only suitable for specific scene.It's not universal.In order to improve the accuracy of 3D point cloud segmentation and sphere identification,the work of this paper is as follows:A point cloud reduction method is proposed by combining the weighted sampling of the outermost points.In order to solve the problem of feature loss caused by noise,a point cloud denoising method based on bilateral filtering optimization is proposed.Under the condition that the point cloud has enough characteristic information.experimental results show that the proposed point cloud reduction method is 7.9 times faster than the traditional sampling method.It removes more invalid point clouds.The P_d and R_dR of the point cloud denoising method proposed in this paper are 96.96%and 92.66%respectively,which are 4.4%and 5%higher than the existing methods.In order to solve the error caused by not considering the relative attributes of adjacent points in the PointNet framework.Based on the multi-scale spatial aggregation framework,a multi-scale point cloud feature segmentation and convolution network framework is proposed.Spatial multi-scale feature extraction module is designed,local features are captured.The sampling is carried out by point cloud simplification and noise reduction.The eight-domain search improved by K-nearest neighbor combined with threshold K was used to get the point features.Multi-directional convolution feature fusion is adopted.The segmentation of point cloud is realized by maximum poolization.On shapanet,the m Io U of classification and object segmentation is 84.1%and 86.7%,which is 3.7%higher than that of PointNet.On S3DIS,the accuracy of segmentation is 81.53%,which is 3.14%higher than that of SAN.A multi-scale sphere feature convolution network framework is proposed for sphere identification which is sensitive to noise.A sphere feature extraction module combining polygonal contour extraction with convex hull processing is designed.Spherical features and point features are fully connected.Local features are extracted from multi-directional space.The multi-scale Sphere Feature Convolution Network framework is validated on the constructed sphere discrimination dataset.The results show that the Io U of the framework is 92.58%,which is 2.9%,3.2% and 15% higher than that of multiscale feature segmentation framework,SAN and PointNet++.In terms of discrimination accuracy,the proposed framework is 94.3%,which is 0.6%,4.8%and 5.9%higher than multi-scale point cloud feature semantic segmentation framework,SAN and PointNet++.The efficiency and feasibility of multi-scale point cloud feature segmentation and multi-scale sphere feature convolution network framework are verified.The related attributes between adjacent points of Point Cloud are obtained by multi-scale Point Cloud feature segmentation model.High-precision segmentation of point clouds is achieved.In the face of the effects of other impurities in complex environments,misidentification can be prevented by a multi-scale sphere recognition network model.Thus,the accuracy of identifying spherical workpiece in industry can be improved.