Research On Key Technologies Of Randomly Sorting Small Workpiece Sorting System Based On Binocular Vision

With the coming of industry 4.0 era,the demand for automation in industrial production is increasing day by day,and machine vision technology is widely used in industrial production.Workpiece sorting is an important part of the common production process,such as parts disassembly and sub assembly,automatic assembly,machine tool processing,etc.The location and recognition of industrial parts is an important direction of machine vision application and research.However,the traditional recognition and location method based on template matching has poor flexibility in application.For small batch and multiple types of workpiece sorting,template design is difficult and costly.Aiming at the common process of automatic disordered sorting and automatic assembly of parts and components in industrial production,this paper studies the key technologies of disordered small-scale parts sorting system based on binocular vision,such as the calibration of binocular vision,the acquisition of depth information of parts,the precise positioning of parts and the identification technology of parts.The main contents are as follows:1.Aiming at the needs of workpiece positioning and identification in the actual industrial production,a binocular vision sorting system is designed and built.The experimental objects are various ball bearings,roller bearings,needle bearings,various gears,screws,nuts and other common workpiece in the industrial production process of sorting and assembly.2.According to the camera model,the relationship between each coordinate system is established,the principle of camera distortion is described,and the basic principle and calibration process of camera calibration are given.In this paper,the standard chessboard calibration board is used to complete the binocular stereo calibration.At the same time,the basic principle of binocular correction and matching,the purpose of stereo correction and the function of binocular matching are introduced.Then,the height measurement of workpiece is completed by using the algorithm of acquiring workpiece depth information based on the parallax principle,which lays the foundation for the subsequent work of workpiece positioning.3.Based on the essential difference between edge points and non edge points in Mathematics-information measure,an edge detection algorithm based on kernel function limit learning machine is proposed,which can extract complete and effective coordinate set of edge points from workpiece image.Through the obtained coordinate set of edge points,the workpiece centroid,geometric principal axis and the minimum external rectangle are calculated to determine the position and orientation of each workpiece.4.Based on the original model network,the structure of darknrt network is improved by using the deep separable convolution network,and the W and H clusters in the data set are clustered by using k-means clustering method.By this way,the prediction frame suitable for workpiece detection is obtained.The labelimg tool is used to label the image set,and the parameters in the data set are normalized to reduce the amount of calculation.The training algorithm model is completed to identify the types of each workpiece.5.The experimental results show that the algorithm designed in this paper can control the result error of the workpiece pose extraction within 1.4mm,the deflection angle error within 0.4o,the average time of each image is 421 ms,and the overall recognition accuracy is more than 99%,which can meet the requirements of industrial production The accuracy and real-time requirements of workpiece positioning and identification.