Design Of Visual Servo System For Manipulator

The problem of manipulator's autonomuous grasping has always been an bottleneck restricting the development of industrial robots and server robots.The recognition and pose estimation of the target is a necessary precondition for manipulator's autonomuous grasping.At present,most industrial robot are still in the teaching stage and cannot make effective judgments on complex environments.It is difficult to achieve effective object recognition especially when objects are densely stacked.In view of the above problems,this paper use the depth camera as the main sensor to reasearch in object recognition and pose estimation problems involved in the robot's grabbing task.At the same time,a complete visual servo system is constructed.The main work of this article includes the following aspects:(1)We design an automatic system for extrinsic parameters calibration based on the marker and Pn P(Perspective-n-Point)algorithm.At the same time,we complete the calibration of manipulator.(2)In this paper,two method of object recognition is proposed.One of them is based on 3D point cloud recognition.The other one is semantic segmentation based on convolutional neural network to achieve the recognition and segmentation of objects.The latter one overcomes some shortcomings of the former and improved the former.(3)Recognition based on point cloud achieve the recognition of object mainly by filtering,segmentation,feature extraction and feature prediction of point cloud.The feature prediction model trained by SVM(Support Vector Machine).(4)Semantic segmentation based on the convolutional neural network compensates for the insufficiency of the original recognition method based on point cloud,so that the object recognition can be more accurate and robust.At the same time,the process of segmentation and recognition are integrated which can simplify the processing of the original method.We improved the original Seg Net's structure which increased the accuracy of segmentation,and it has been verified on public data sets.(5)The 6-DOF pose estimated by using ICP(Iterative Closet Point)algorithm,and the pose estimation's mean error is 8.89 mm.(6)We package the algorithms above into a complete visual servo system,which enables automatic sorting of the manipulator and confirms the effectiveness of the entire method.The visual servo system designed in this paper achieve the objects' recognition and 6-DOF pose estimation by point cloud processing and semantic segmentation.It has high recognition and pose estimation accuracy,solves the problem when objects placed densely,lays a good foundation for manipulator's grabing and imporves the intelligence of the robot.