Feature Segmentation And Location Recognition Based On 3D Point Cloud

It's increasingly commonplace to use roborts for intelligent detection,sorting and other tasks in manufacturing production lines.In the field of machine vision,these tasks can be divided into scene understanding,object recognition and pose estimation.At present,due to the maturity of 3D imaging technology and the updating of 3D scanning equipment,machine vision algorithm based on 3D data has become a research hotspot of scholars.Therefore,based on the three-dimensional point cloud data,this paper mainly studies the point cloud segmentation algorithm based on supervoxel spectrum clustering and the improved algorithm of pose estimation based on point pair feature descriptors and ICP registration.The main work of this paper is as follows:(1)A point cloud reduction method,which preserves the details and boundary features is studied.Firstly,the acquisition technology,neighborhood search technology and normal estimation technology of point cloud are analyzed.Moreover,aiming at the problem of massive point cloud data reduction,a simplification method is proposed to distinguish the point cloud boundary with weighted joint force,and simplify the internal point based on Gauss sphere mapping and mean shift clustering.Compared with the common grid reduction and curvature reduction methods,this method is more effective in showing details and boundaries.(2)A method of spectral clustering segmentation based on supervoxel is studied.Firstly,the basic theory of spectral clustering is described,including Laplacian matrix,spectral map partition strategy,etc.Secondly,based on the perturbation theorem of Laplacian matrix,a method of automatically determining the number of spectral clustering is studied.In addition,the fusion process of supervoxels,including voxel partition and similarity measurement is introduced.On this basis,KFCM algorithm is used to map supervoxels to high-dimensional feature space and cluster them to get their corresponding membership matrix.According to the membership matrix,similarity measurement method of supervoxels is obtained.Finally,spectral clustering algorithm is used to complete the fusion of supervoxels.(3)The improved algorithm of pose estimation based on PPF feature and ICP algorithm is studied.This method optimizes the calculation of rotation angle of coordinate transformation based on PPF feature.In order to eliminate the ambiguity of normal,a method based on minimum spanning tree is proposed to adjust the orientation of normal.In the process of initial pose estimation,the role of FPFH feature descriptors in describing local features is considered to improve the accuracy of reference point matching.At the same time,in the process of pose optimization,the traditional target of point-to-point distance deviation is replaced by the function of point-to-surface distance deviation,and due to the normal constraint,the accuracy of pose optimization of ICP algorithm is improved.In the process of algorithm verification,this paper evaluates the results of pose estimation,occlusion analysis and noise robustness test of multiple instances,and proves the feasibility and stability of this algorithm.