Research On Positioning And Grasping Of 3D Parts Based On Monocular Vision

In the traditional sorting process,the sorting task is generally manual sorting method,and this sorting work is monotonous and lacks creativity.With the development of industrial automation,machines can be used to replace humans to solve this monotonous work.The sorting task is the typical application of the combination of machine vision and manipulator.Aiming at the problem of monocular recognition and grabbing of non-textured reflective objects on the surface,this paper proposes a monocular vision sorting system.This article focuses on the hardware selection of monocular vision platform,platform construction,vision system calibration,image preprocessing,and target matching algorithm.The main contents are as follows:First,a monocular visual sorting platform was designed.For an excellent vision system,the system hardware must achieve perfect coordination.Therefore,the hardware selection of the vision sorting platform is carried out according to the field of view,the viewing distance,and the working environment.And for non-textured reflective objects with the help of Halcon,its image preprocessing makes the edge features more obvious.Secondly,the problems of camera calibration and hand-eye calibration are studied.The dot calibration board and the checkerboard calibration board were used to carry out camera calibration experiments,and the influence of different calibration boards on the calibration was analyzed.The characteristics of the two hand-eye calibration installation methods are studied,and then the selected eye-to-hand model is used to complete the hand-eye calibration with the help of Halcon to obtain the positional relationship between the manipulator and the camera.Again,the pose estimation algorithm is studied.The three-dimensional matching algorithm based on CAD template is studied,and it is found that the algorithm is more sensitive to the background and is prone to mismatch.Aiming at this defect,a method of combining a 3D matching algorithm based on a CAD template and a deep learning algorithm is proposed to reduce the interference of the image background through the deep learning algorithm.Among them,the deep learning algorithm uses Yolo-lite’s design ideas to design and train the Yolo-lite2 algorithm to obtain reliable results.Finally,a comparative experiment on the joint algorithm solves the background interference problem and speeds up the matching speed.Finally,the experimental verification of the visual sorting platform was completed.The software part of the system is designed in modules and the software interface is written.The multi-posture grasping of a single target object and the sorting of multiple types of target objects were tested respectively,and the error analysis of the sorting results was carried out.The results show that it can complete the grasping of various target postures and sorting tasks of multi-target objects,which meets the design requirements of the system.