Research On Object Recognition And Grasp Based On Faster-RCNN

Robot grasping technology has always been a hot issue in all walks of life,from the initial through the robot teaching programming to achieve the fixed position and fixed types of workpieces to grasp,and then gradually through the visual system to locate the workpiece.Therefore,robot recognition before grasping has always been the research object of frontier technology.Traditional recognition and grabbing technology,especially in industry,is mainly through machine vision,that is,the image processing technology is used to obtain the contour size of the workpiece,and then the deflection angle between the workpiece normal and the reference normal is obtained to obtain the posture.This method requires fixing the type and shape of the workpiece.If abnormal workpiece suddenly appears,the system may produce errors.In recent years,with the rapid development of deep learning technology,the application of deep learning technology in recognition and grasping of robots can greatly improve the intelligence of robots and reduce the workload of maintenance personnel.In this context,this paper combines image recognition technology with point cloud to complete the recognition and location of the workpiece,and then complete the task of recognition and grasp.This paper mainly completes object recognition based on Faster-RCNN,and then obtains the pose of the target object through point cloud.The overall idea is to obtain the color image and depth image of the target object through Kinect camera.After that,the color image is sent to the trained recognition network for object recognition,and the range of the object in the image is framed.At this time,the depth data of the same location and the same range size are extracted from the depth image,and the depth data are transformed into point cloud data and registered with the model point cloud of the corresponding object that has been calibrated.The registration data obtained is the position and attitude information of the object to be grabbed relative to the reference model point cloud.The pose information is transmitted to the robot,and the robot makes the corresponding trajectory planning according to the relevant data to complete the grasp.In the training part of the recognition network model,the identification network is constructed by reforming the ZFNet network.Preserve the feature extraction part before the full connection layer of ZFNet network.Using the trained feature extraction parameters can simplify the training time,and then create two full connection layers according to their own data sets.At the same time,Faster-RCNN uses RPN network to generate extraction frames.Therefore,RPN networks also need to be trained through data sets.RPN network will output two sets of results according to the feature map of feature extraction part.One is used to distinguish the foreground or background of the feature in the image,and the other is used to return the extraction box.The registration part of point cloud consists of two parts.The first part is the production of the reference point cloud model.The second part is to extract the point cloud from the center of the view and make the corresponding labels at the same time.At this time,the position of the robot arm should be fixed,and so as to complete the binding of the benchmark model and the robot arm.After that,the object to be grabbed is placed arbitrarily in any position in the camera window,and the point cloud is extracted by the results of the previous recognition part.Then,the position and attitude information of the object to be grabbed relative to the reference model is obtained by ICP point cloud registration method.The part of robot grasping.Through the processing of the first two parts,the pose information data of the object to be grabbed is obtained.The trajectory planning of the robot is carried out according to the corresponding position and posture data to obtain the transition points from the starting point to the terminating point.Accurate movement of robot manipulator is realized.At the same time,this part also needs to apply socket communication protocol,which is used to send pose data after computer processing to the robot controller.At the same time,it receives the feedback from the controller to the computer.This paper presents objects recognition and grabbing method based on Faster-RCNN.Firstly,the type of workpiece and its position in the image are identified,then registration is completed by using the point cloud data formed by the depth data of the corresponding position,and the position and attitude information obtained is transmitted to the robot to complete the corresponding grasping task.This method integrates recognition,positioning,motion planning and other technologies to achieve the goal task.It can give full play to their respective technical advantages and characteristics,and enhance the robustness and stability of the system.