Research On Naturally Stacked Objects Recognition And Localization Based On Fringe Projection

In recent years,digital fringe projection profilometry is widely applied in various fields,such as reverse engineering,dimension measurement,intraoral scanning,robot navigation,etc.Compared with other technologies,fringe projection profilometry has the advantages of low cost,fast measurement speed,large field of view and high spatial resolution,so it is favored by the market.As a fundmental research area in computer vision,object recognition and localization technology has numerous applications,such as automatic assembly,biometric analysis and medical treatment,intelligent surveillance,remote sensing,robot manipulation,etc.This paper mainly uses the 3D measurement technology based on fringe projection to obtain point cloud data of the scene,and uses point cloud processing algorithms to realize the recognition and localization of naturally stacked objects.Detailed research work is summarized as follows:(1)Based on the principle of fringe projection 3D measurement,the methods of obtaining relative phase,absolute phase and height are expounded.Among the absolute phase acquisition methods,multiple temporal phase unwrapping methods are compared.The mapping relationship between height and phase is established.A 3D measurement system is set up to obtain point cloud data of the scene.(2)The source of the phase error is analyzed,and the phase error caused by the nonlinear gamma of projector is analyzed mathematically.Passive projector nonlinear gamma compensation method for phase error is studied,and an improved method for construting a look up table is proposed,reducing phase error by 91.37%.The mathematical model of camera and projector calibration is studied.On the basis of the inverse camera model,an improved projector calibration algorithm is proposed,using a homemade dot calibration board,reducing the number of images projected by projector,making calibration operation simpler and faster.The absolute measurement accuracy is about 0.1mm,and the relative accuracy is 1:5860.(3)Passthrough filter,voxel grid filter,statistical outlier removal and radius outlier removal are analyzed,redundant points and outliers in raw point cloud data are removed and the density is reduced.Various point cloud segmentation algorithms are analyzed in brief,and Euclidean clustering segmentation is expounded.As the existing segmentation algorithms cannot realize the complete segmentation of eyebolt for different kinds of poses,an improved Euclidean clustering segmentation algorithm is proposed and resolve this problem.(4)The algorithm flow of 3D object recognition is designed.Various feature description and extraction algorithms of point cloud are compared,and the FPFH descriptor is selected.The matching principle based on SAC-IA coarse registration and ICP fine registration is expounded,and the combination of two algorithms is used to realize the recognition and localization of naturally stacked objects.The recognition time of a single object is about 1.5s.