Detection, Pose Estimation And Grabbing Of Steel Plates Based On Robot Vision

Three-dimensional object detection and pose estimation is the basis of robot perception environment.It gives the robot a certain degree of "intelligence" to realize the tasks of loading,unloading,assembling and handling without human interference.However,when the robot is in an unstructured environment,the disorderly staggered stacking and occlusion between the same or different objects bring great difficulties to object detection and pose estimation.Therefore,a method of three-dimensional object detection,pose estimation and grasping is studied in this paper,which can enable the robot to complete the task under the above complex conditions.The practical application of steel plate grasping is also studied.The main work of this paper includes the following aspects:1)Aiming at the lack of small object detection performance in SSD network,an improved feature fusion algorithm for small object detection is proposed,which effectively improves the detection performance of small objects.By using deconvolution,the high-level features with strong semantic information are added to the low-level feature layer,and the skip connection is used to accelerate network training to improve the ability of object detection.At each feature level fusion,an adaptive weighted connection is proposed based on the feature of neural network automatic learning.2)In order to reduce the results of repeated recognition and false recognition in template matching,a template matching algorithm for region of interest of target object is proposed.In this paper,an improved SSD is used to obtain the region of interest of an object,and only template matching is used in the region of interest to reduce the false recognition rate.For further eliminate the local false recognition of objects,a template clustering method based on similar spatial location is proposed by improving the template clustering method.3)For the pose estimation methods of planar objects,two different methods are proposed according to different requirements.One is a rough pose estimation method designed according to plane features,which can meet the needs of fast but low precision grasping.The other is a fine pose estimation method based on template matching,which is suitable for planar or three-dimensional objects.In addition,the inaccurate result of pose estimation is further eliminated by using pose verification strategy.4)Aiming at the problems of camera accuracy and field of view,the methods of two detection and two pose estimation are proposed.The purpose of the first detection and rough estimation of object pose is to make the end of the robot move directly above the object,so as to minimize the error of point cloud generated by the camera.The purpose of the second detection and fine estimation is to minimize the estimated pose error.5)Hardware platform is built with depth camera,color camera,magnetic end effector and UR3/5 robot.Software framework is built with ROS system,Opencv image processing library,PCL point cloud processing library,MoveIt motion planning library and TensorFlow neural network library.The validity and feasibility of the whole system are verified in the experiment scene of grasping objects stacked in chaos.