Research On Visual Inspection Method Of Safe Operation System Of Industrial Robot

With the transformation and upgrading of China's intelligent manufacturing industry,the demand for industrial robots has increased dramatically.Human machine cooperation is an important feature of robot development.All kinds of robots and people work together in the workshop,so the workers in a weak position have relatively large potential safety hazards.According to the U.S.labor office,4585 workers have been killed in robot related accidents since 1984.In order to ensure the safety of operators in the human-machine cooperation environment,this paper studies the vision detection technology in the safe operation system of industrial robots.In this article,the human body information detection in the process of human and fixed industrial robot collaborative work is taken as the research object.A three-dimensional human body image visual detection system based on three Kinect 2.0 depth cameras is constructed.The hardware part of the system is mainly composed of three Kinect 2.0 cameras and a computer.The software environment is mainly based on Windows 10 system using visual studio 2017 compiler and configuration installation pcl 1.8.1 development environment as the development platform.The main work is as follows:Firstly,two kinds of errors that plane error and depth error of Kinect 2.0 camera are corrected.Two dimensional imaging plane error is corrected by chessboard calibration.For the depth error of depth image,a spatial registration model of composite function wavelet neural network sequential limit learning machine(CFWNN-OSELM)is proposed to uniformly correct the nonlinear depth offset and systematic depth overestimation error caused by Kinect 2.0 camera measurement principle.The algorithm proposed in this article can not only achieve online incremental fast learning,but also has stronger generalization ability.Experimental results show that the algorithm can recover the depth data better in real scenes.It lays the foundation for the high-quality point cloud data in the field of view obtained by the Kinect 2.0 camera.Then,on the basis of single Kinect 2.0 correction results,the automatic segmentation method of point cloud data of human body is studied.In this article,we use Kinect 2.0 bone data to calculate the center coordinate of human body in three-dimensional space as the pass through filter threshold to automatically segment the point cloud data of human body from the original point cloud data.The simplification of the original point cloud data is studied.Firstly,the voxel grid method is used to automatically determine the voxel size of the original point cloud data in the process of downsampling by using the human 3D space center coordinates calculated based on Kinect 2.0 bone data.Then,based on Kinect 2.0 bone data,the center coordinate of human body in three-dimensional space is calculated as the threshold value of pass through filter to automatically segment the human body target cloud data from the point cloud data after de sampling.Finally,a series of discrete points are removed from the point cloud data of human body which is successfully segmented by automatic positioning by using the statistical outlier removal method.The experimental results show that this method can distinguish the human target model in complex scenes,so as to obtain the flat point cloud data quickly and accurately.Thirdly,aiming at the problems of low recognition accuracy and low real-time in the existing human motion recognition under the complex background,a registration model recognition algorithm of human posture limit learning machine(ELM)is proposed,which uses Kinect 2.0 to measure and calculate the joint association angle.Through Kinect 2.0,the spatial coordinates of human joints are obtained.Then,three-point method is used to calculate the relative angle of human joints,and then the movement recognition test is carried out through the registration model of human body posture limit learning machine(ELM).This method can identify more postures by defining the joint related angles of more target postures to establish the elm registration model of human postures.Finally,aiming at the registration and splicing of point cloud data involved in the three Kinect 2.0-based 3D human body image visual detection system,a two-step method of coarse registration and fine registration is proposed to register the point cloud data obtained by each camera.Due to the commonly used registration algorithms such as 3DSC,PFH,FPFH and NDT,the coarse matching effect is better when the difference between the two images is not big,but for this paper,because the common parts of the two target point cloud data images that need to be registered are less,there are some differences between the point cloud data images,and the effect is not ideal.In order to better meet the needs of this paper,the first step is to use basketball as the target,calibrate the external parameters of three cameras in one to one coordinate system,that is coarse registration;the second step,based on the coarse registration,precisely register the point cloud data obtained by the three cameras through ICP algorithm,so as to achieve the goal of accurate registration of complete and smooth 3D point cloud data model of human body.