Research On Target Recognition And Location Of Visual Manipulator

In recent years,robotic arm technology has become more and more widely used in various fields.However,the existing orchard picking robot arm and industrial production robot arm have many disadvantages.For example,some robotic arms are equipped with relatively backward vision systems,or even have no applied vision systems.These problems hinder the development and application of robotic arms.In practical applications,how to enhance the recognition and positioning capabilities of the robotic arm system and the ability to complete complex tasks have become the key issues for improving the robotic arm system.In view of the lack of recognition and positioning capabilities of the robotic arm system,this thesis conducts the following research:At present,the traditional combination of the vision system and the robot arm system is to fix the camera at one end for identification and positioning,and the grasping task is completed by the robot arm at the other end.This design method has limitations,and the object to be grasped must be fixed according to the position of the camera.In order to improve this disadvantage,this thesis proposes a recognition and positioning system that fixes the depth camera to the end of the robotic arm.This construction method can be used for multiple and multi-angle recognition and positioning according to the target placement angle,increase the generalization ability and practicality of the system.In terms of recognition,deep learning algorithms are more accurate than traditional algorithms based on feature matching.Therefore,this thesis proposes to apply the YOLO v3 algorithm in deep learning to the system,and to improve the problem of the algorithm's inaccurate recognition of small targets and inaccurate recognition frame regression.In this thesis,the number of layers of the YOLO v3 basic network is reduced,and a multi-scale feature extraction structure is added.Three different sets of convolution kernels with the size of(1×1),(1×1 +3×3),(1×1 +3×3 +3×3),and are used to extract features and stack them together.Finally,for the tedious problem of setting the anchor box in the YOLO v3 algorithm,the Mean-shift algorithm was adopted to improve the size and number of the anchor box.In terms of positioning,the "eye-in-hand" calibration method is used to obtain the three-dimensional world coordinates of the target.At the same time,the grasping task uses the linear interpolation algorithm in Cartesian space for motion path planning.Finally,the recognition and positioning accuracy of the visual robotic arm system is above 95%,and the average grasping accuracy rate is 90%,which is improved compared to all aspects of the traditional robotic arm system.