Research On Parts Recycling And Grasping Based On Convolutional Neural Network

With the continuous upgrading of modern industrial technology,the overall production scale of my country’s machinery manufacturing industry has ranked among the top in the world.Correspondingly,the use of mechanical parts is increasing day by day,and the recycling and reuse of parts is of great significance to social resources and the natural environment.At present,in parts factories,robot intelligent grasping technology gradually replaces manual labor in the fields of production,manufacturing and parts recycling by virtue of its high efficiency and accuracy,greatly improving work efficiency and factory intelligence.However,most of the researches on parts grabbing and detection by domestic robots still remain in the traditional feature matching and edge detection,and the recognition accuracy and speed are slow,and the generalization ability is weak.In the actual parts recycling factory,the segmentation accuracy of scattered parts is low,which makes it impossible to accurately identify and grasp.In view of the above problems,this paper first calibrates the robot grasping vision system.Then,a variety of target detection algorithms under convolutional neural network are studied and compared.The grasping mathematical model of the robot end effector of the three-fingered manipulator is established.In order to obtain the grasping parameters,an improved grasping detection method based on convolutional neural network algorithm Mask R-CNN is proposed and experimentally verified.The research mainly includes the following aspects :(1)The calibration of robot vision system is studied,and the conversion relationship between the coordinate systems in the vision system is obtained.The mathematical principles of camera internal and external parameter calibration are studied,and the calibration accuracy is improved by using the multi-distortion iteration method.The robot hand-eye calibration is carried out,and the image pixel coordinate system is converted to the robot coordinate system.(2)Researched the basic theory of convolutional neural network,analyzed and compared the basic architecture composition and characteristics of multiple target detection networks based on the candidate region series.Focused on the improvement of Mask R-CNN in this series,and obtained Mask R-CNN has a conclusion on the superiority of parts segmentation in complex environments.(3)In order to realize the grasping operation of the robot,a mathematical model based on the end-effector grasping of the experimental robot was first established,and the problem of grasping was represented by parameterization.And then put forward a kind of fetching detection based on Mask R-CNN network improvement method.The residual network used for feature extraction into balance and pyramid thinking before replacing the traditional characteristics of the pyramid.The loss function are analyzed and improved.In the output layer increased the branch Angle prediction,and combined with the depth of the images provided by the laser line scan camera coordinate parameters fetching model are obtained.Finally,the corresponding data set is established for this method,and the network training and part inspection verification experiments are carried out.(4)The robot grasping detection platform is built and the experimental verification of the grasping method is carried out.Firstly,the components of the robot grasping detection platform are introduced,and the forward and inverse kinematics analysis of the experimental robot is carried out.The robot grasping posture can be planned when the grasping parameters of the robot end-effector are known.Finally,on the basis of the above research,the robot autonomous grasping experiment is carried out,and the experimental results are analyzed to verify the reliability of the grasping method proposed in this topic.