Research On Spool Pose Estimation And Grasping Based On Deep Residual Network

In the manufacturing process,improving the automation of the sorting process of industrial parts is of great significance to improving the efficiency of production operations.Compared with the traditional orderly sorting,the vision-based disorderly grasping needs to classify the goods in advance,which has higher flexibility.Therefore,this paper uses vision to guide the robot,and proposes a valve core part pose estimation algorithm based on a deep residual network to realize the grasping of the actual valve core part.The work content is as follows:(1)Design the overall scheme including the data set module,the pose estimation module and the robot control module.The data set module uses the Aruco positioning method to obtain the three-dimensional data set of the spool for the training of the pose estimation module;the pose estimation module combines deep learning technology and Pn P algorithm to calculate the relative pose relationship between the camera and the spool;the robot control module Developed based on the ROS platform to realize functions such as robot control,physical simulation,trajectory planning,and hand-eye calibration;in addition,the checkerboard calibration method is used to calibrate the internal parameters of the camera for the calculation of the pose estimation module,and the Tsai-Lenz algorithm is used for hand-eye calibration Calculate the relative relationship between the camera and the robot base coordinates;finally,use ADD(Average Distance of Model Points for 3D object detection)and 2D-Proj(2D-projection error)to evaluate the pose accuracy.(2)Use the method of fusing the 3D model to generate the spool data set to improve the accuracy of the data set model.According to the Aruco mark in the scanned picture,the Aruco code 6D pose of the scene is obtained through Open CV processing;and the Poisson sampling method is used to sample the surface of the 3D model of the valve core part to obtain high-precision point cloud data of the valve core model.And use the ICP method to register with the scanned rough point cloud in the scene to improve the accuracy of the dataset;finally use Blender to render the dataset synthetically to expand the number of datasets.(3)Multi-scale improvements are made to the PVNet network algorithm to improve the recognition accuracy of valve core parts.On the basis of its backbone network,a bottleneck layer unit based on group convolution is designed to improve the multi-scale performance of the network model;using the improved network algorithm,regression obtains the unit vector of each pixel pointing to the key point,through the vector field The voting algorithm obtains the 2D key point positions;then the 3D key points obtained by the FPS(Furthest Point Sampling)algorithm are matched and mapped with the 2D and 3D key points using the Pn P algorithm,and the relative pose relationship of the parts is obtained by solving.(4)Build an experimental platform,the hardware part designs a robot grasping platform based on vision,the software part uses the produced data set for training,and obtains the network learning model of the valve core parts;according to the determined pose evaluation method,the grasping experiment is carried out According to the analysis of the results,the experimental results show that the success rate of valve core parts capture is 92%.Compared with the network algorithm before the improvement,the multi-scale improvement improves the accuracy by about 20%.