Research On The Robotic Grabbing Technology Of Scattered Parts Based On 3D Vision

As the important part of industrial automation,robots are needed to provide strong support for intelligent manufacturing.Machine vision can improve the vlexibility,the maneuverability and the intelligence of robots.Currently,visual information used for capturing object generally comes from two-dimensional images and three-dimensional(3D)point clouds.Compared with the two-dimensional images,the 3D point clouds include the depth information of object,which makes the perception ability of robots more powerful.Although 3D vision currently plays a role in certain scenes,it still cannot fully meet the requirements of social development for production automation.In order to solve the problems of identifying and locating scattered objects,RealSense is used to obtain target point cloud.The posture of object is obtained through point cloud segmentation and point cloud matching.Then the grasping path is studied and an experiment is conducted to verify the proposed method.Firstly,RealSense depth camera is used to obtain the point cloud of scattered parts.For the jump phenomenon and the holes phenomenon in depth images,Kalman filtering and joint bilateral filtering are used to improve the quality of original 3D point cloud.For the problems of large number and noise of 3D point clouds,pass-through filtering,voxel sampling and discrete points removing are used to extract point clouds,reduce the number of points and subsequent processing time.Then the point cloud segmentation method based on super voxel is applied to separate the scattered overlapping objects.For the point cloud whose color is same,original color criterion is abandoned,while boundary constraints are added to the segmentation process,which can improve the accuracy of segmentation as well as successfully obtain the target object from multiple scattered objects.The method of part recognition based on fast point feature histogram(FPFH)and the method of part pose determination based on point cloud registration are proposed.For the selected target object,the quick point feature histogram is extracted and compared with the template object.The type of target object is judged according to the similarity between the two objects.Then an improved ICP algorithm is proposed to optimize the parameters of pose transformation matrix for obtaining the pose matrix between the point clouds of the template object and the grabbed object.Finally,the methods of grasping path planning and grasping stability judgment are proposed.Force closure criterion is used to design a reasonable grasping orientation.Hierarchical bounding box method is used to detect whether there is a collision between the manipulator claw and the target object.An improved PRM algorithm is proposed to plan and optimize a collision-free grasping path.