Research On 3D Pose Recognition System For Robots Grasping Objects

In recent years,in order to improve production efficiency and reduce labor costs,a large number of warehousing and logistics manufacturers have requested to increase the corresponding automatic sorting line to achieve automatic sorting,automatic detection of goods and so on.Among them,the realization of automatic object recognition and grasping system is not only of theoretical significance,but also of high practical value because of its wide range of application,and has become the focus of many experts and scholars at home and abroad.In this paper,the related technologies such as machine vision on-line detection,image processing and robot system integration are deeply studied,and a set of shelf object automatic recognition and grasping system is designed and developed by taking the shelf object pose recognition as an application case.The main research contents are as follows:Firstly,the hardware system is designed to meet the needs of the system,and a complete experimental platform is designed and built.Six common object models are selected as the experimental verification materials for the experiment,and a platform for loading objects is established.By comparing and analyzing the advantages and disadvantages of the two robots in the laboratory,the UR5 robot for grasping action is selected for the whole system.At the same time,the characteristics of the three-dimensional camera on the market are analyzed comprehensively.Real Sense F200 is selected as the experimental camera,and the depth camera and the manipulator are combined to fix the manipulator on the robot.At the end,the "hand-eye" is combined as a whole,providing forward kinematic support for determining the attitude of the object on the shelf.Secondly,the advantages and disadvantages of ROS and Player are compared and analyzed.According to the characteristics of this experiment,ROS is selected as the overall framework of the system.The position of the image is determined more accurately by using SURF algorithm.Compared with traditional algorithm and neural network,the characteristics of convolution neural network and full convolution neural network are analyzed.A more robust full convolution neural network is selected to segment the object in the whole scene.After binarization,the template image of the object is obtained,and the whole point cloud of the object is obtained by "multiplying"the obtained depth image with the template image.Then,after background removal and denoising,the PC A pose of the point cloud is solved,and the PC A pose is adjusted by two steps of ICP,which makes the PCA position more accurate.Finally,the relative position relationship between the camera and the manipulator is determined by hand-eye calibration,which provides a basis for the transformation of the pose of the object in the camera space into the world coordinate system;the relative position relationship between multiple positions and poses is obtained by using the feature images with more feature points,and the spatial point cloud is provided.The fusion provides forward kinematics relationship,and transmits the acquired object posture to the robot motion system through the ROS framework.The manipulator reaches the position of the object and performs the grasping action of the manipulator.