Research And Application Of Casting Recognition And Pose Estimation Based On Machine Vision

A highly flexible casting production line is a necessary condition for producing high-precision and complex castings,which often has many processes,complex operation processes,and frequent task switching.The transfer of castings between multiple processes is an important factor affecting the efficiency of the production line.Generally speaking,each process of flexible production line often produces a variety of different parts,which are placed in the box in a random and disorderly way.At present,the transfer of parts between various processes depends on labor,which leads to the inconsistency of production rhythm and low operation efficiency.Aiming at the actual industrial scene of disorderly grasping of castings,this paper studies the algorithms of casting recognition and pose estimation based on machine vision.The main contents are as follows:(1)The casting single target pose estimation algorithm is analyzed,the speed and accuracy performance of different point cloud feature descriptors in pose estimation are compared,and the reasons affecting the running speed of the algorithm are analyzed.Through comparative experiments,the feature descriptors suitable for casting disorderly grasping scene are determined,and the accurate pose estimation based on SAC-IA coarse registration and ICP fine registration algorithm is realized.(2)Aiming at the problem of too many high-precision point clouds and too long pose estimation time.A down-sampling method with balanced significance and uniformity of point cloud is proposed.The visual significance of point cloud is measured by combining a variety of geometric features and fused with the uniformity of point cloud to realize lossless down-sampling of point cloud.Three dimensional reconstruction and registration experiments were carried out on three kinds of castings,and three-dimensional reconstruction experiments were carried out on more complex parts and free-form surface parts.Experiments show that the algorithm has good robustness,and the root mean square error(RMSE)of3D reconstruction of this method is reduced by 40%,32% and 19%respectively compared with K-means clustering,grid averaging and graph based down-sampling methods at a down-sampling rate of 10%.In addition,the time-consuming of object pose estimation is reduced by 80%.(3)Aiming at the pose estimation problem in multi-target disorderly stacking scene,a method of 2D and 3D vision fusion is proposed,the casting recognition and positioning data set is established,and the accurate extraction of casting ROI region is realized.The mapping relationship between 2D image and 3D point cloud space is established,and the fusion of 2D extraction results and 3D point cloud data is realized.The bullet physical engine is used to simulate the point cloud data,and the algorithm is verified on the simulation data.(4)A casting recognition and pose estimation system in complex scenes is developed.Taking the Jaka Zu 7 industrial manipulator as the carrier,the ALSON structured light camera is used to collect RGB images and point cloud data,and the TCP calibration of the tool coordinate system at the end of the manipulator and the hand eye calibration between the camera and the manipulator are completed.The experiments of disorderly grasping in real scenes verify the effectiveness of the recognition and pose estimation algorithms.The experimental results show that the average success rate of the algorithm can reach 96.88% for three scenes with scattered castings.Figures 58,tables 18,references 79...