Research On Machine Vision Of Mental Workpiece Sorting System

With the "Industry 4.0" in Germany and the "Advanced Manufacturing Partnership Program" in the United States have been proposed,more and more countries are committed to the realization of industrial intelligent manufacturing.As the basis of intelligent manufacturing,the machine vision technology is added into the workpiece sorting system,which not only improves the degree of system intelligence,but also broadens the application scene of the system.The development of industrial intelligent manufacturing needs to improve the workpiece recognition rate and the accuracy of workpiece pose estimation in the sorting system.This paper takes metal workpieces,the most used in industrial sorting,as an example to conduct an in-depth study on improving the efficiency and recognition rate of the workpiece sorting system.The main research contents of this paper are as follows:(1)Aiming at the problem that the calibration results of traditional hand-eye calibration algorithm have large errors in practical applications,an optimization algorithm based on corner preprocessing is proposed.The checkerboard is used as a calibration board to extract the corners,and the reprojection error of each corner is calculated.Then,according to the adaptive reprojection error threshold,the corner coordinates are selected to calculate the hand-eye calibration results.Finally,the coordinates of the four internal vertices of the calibration board measured by the CGXi G6 robot are compared with the coordinates calculated by using the calibration results to verify the accuracy of the algorithm.The experimental results show that compared with the original hand-eye calibration algorithm,the proposed algorithm has the average reprojection error of corner extraction reduced by 18.98% on average,and the coordinate calculation accuracy increased by 33.18% on average.(2)In industrial sorting scenarios,it is hard to distinguish workpieces with high similarity with low-resolution 3D point cloud data,and the point cloud processing speed will decrease with the increase in the number of workpieces.Aiming at the above problems,a point cloud classification and segmentation algorithm based on YOLOv5 is proposed.The image data of the target workpieces are collected to build the data set and train the YOLOv5 target detection model.Then,the trained model is used to classify the target workpieces,the prediction box coordinates in the model prediction results are mapped to the point cloud coordinate system,and the conditional filtering algorithm is used to segment the point cloud of the target workpiece from the original point cloud.Finally,the plane fitting algorithm based on Random Sample Consensus(RANSAC)is used to remove the noise of the placement plane in the segmented point cloud,and the target workpieces’ category and point cloud data are obtained,so as to complete the classification and segmentation of high similarity workpiece point cloud.The experimental results show that the proposed algorithm can accurately obtain the category and point cloud data of the target workpieces,reduce the point cloud data involved in the subsequent point cloud processing,and meet the requirements of industrial applications.(3)When using Fast Point Feature Histogram(FPFH)to extract feature descriptors to complete the point cloud registration task,the neighborhood radius needs to be set in advance,and the inappropriate neighborhood radius will lead to long time-consuming,and low accuracy of point cloud registration.To solve the above problems,a point cloud registration algorithm based on FPFH feature extraction and adaptive neighborhood radius is proposed.The ISS3 D algorithm with an adaptive search radius is used to extract the key points of the original point cloud,and the FPFH feature extraction algorithm is used to extract the feature descriptors of the key points.Then,the RANSAC algorithm with geometric constraints is used to calculate the registration matrix of the key points as the initial pose of the original Point cloud registration,and the Iterative Closest Point(ICP)algorithm is used to obtain the accurate registration result.Finally,the average distance of the point cloud corresponding to the neighborhood radius of different point clouds when the registration performance is optimal is calculated,and the multilayer perceptron is used to fit the mapping curve between the average distance of the point cloud and the neighborhood radius to realize the automatic selection of the neighborhood radius.The experimental results show that compared with the four registration algorithms in other papers,the registration accuracy is improved by 52.06%,59.91%,83.85%,and 20.06%,and the registration speed is increased by 41.38%,6.67%,67.79%,and 74.45%,respectively.(4)In order to realize the designed metal workpiece sorting system based on machine vision,the CGXi G6 robot and Tuyang FS820-E1 camera are used to build an experimental platform.The hand-eye calibration optimization algorithm based on corner pretreatment is utilized to calibrate the camera,the point cloud classification and segmentation algorithm based on YOLOv5 is adopted to obtain the point cloud and category of the metal workpiece,the point cloud registration algorithm based on the adaptive neighborhood radius FPFH feature extraction is used for registration,and the pose of each metal workpiece in the robot base coordinate system is estimated respectively.The experimental results show that the designed sorting system can identify metal workpieces with high similarity,and the sorting speed and accuracy meet the basic requirements of industrial applications.