Study On The Visual Sorting Method Of Industrial Robot Based On Deep Learning

With the coming of industry 4.0,the industrial robot plays an increasingly important role in intelligent factories.On the one hand,the workers are replaced by the industrial robots to perform the repetitive work,and the workers are liberated from the boring work environment.On the other hand,it saves production cost and improves the production efficiency.As a result,an increasing number of manufacturing industries begin to use the industrial robots for manufacturing.On the traditional industrial production line,the sorting task is carried out manually,which is not inefficient,but also costly.Therefore,it is the trend of industrial automation to apply machine vision technology to the sorting task of industrial robot.At present,the identification and positioning are the main problems to hinder the popularization of industrial robot in sorting tasks to the work piece that is of relatively complex shape in the cluttered order.This paper introduces the deep learning technology to realize the detection and accurate positioning of the work piece in terms of the problems that traditional visual algorithm are difficult to identify and locate the work piece with different standing and dense placement.In this paper,it uses the target detection and positioning,camera calibration,network communication and other contents as entry points and carries out the study on the visual sorting method of robot based on deep learning.A total of three core issues are involved as: The first is the calibration of camera,the second is the work piece detection and key point positioning problem based on deep learning technology,and the third is to build the robot sorting system platform for experiment.The main contents are as follows:This paper is aimed to study the camera imaging geometry model,establish the linear and nonlinear model of camera and deduce the principle of camera calibration.Also,a checkerboard calibration method is designed according to two models,and furthermore,the one-to-one corresponding relation between image pixel points and space objects on the robot working platform is established.This paper proposes a CNN model for the detection of work piece on the industrial site assembly line.The model is improved on the basis of the classical Faster R-CNN model,and unified such three processes as image segmentation,target feature extraction and classification.The improved model is named PDN.In this paper,one of the innovation points is to study the candidate region extraction technology of regional proposed network(RPN)and proposes an improved model to improve the performance of the RPN network.The improved area proposed network area is named ZRPN.To the problem of work piece positioning,it proposes another CNN model,which is named KPPN.The model uses European loss as the loss function,to access the pyramid pooling layer at the end of the last convolutional layer so as that the KPPN model can accept any size of images.The KPPN model can directly take the original work piece image as input,without any image preprocessing operation.Output is set with four values,which represent the two key coordinates of the work piece.With establishment of the visual sorting system of industrial robots,compilation of human-computer interaction software and a lot of real scene experiment verification,it is proved that the model presented in the paper is of certain robustness and effectiveness in many scenarios,such as occlusion,illumination change and camera height change.