| With the development of robot technology,robot has been widely applied to the fields closely related to human's production and life,and play an important role.Grasping is an important capability of human-machine coordination for robots in both service and industrial scenes.Obtaining an accurate grasping detection result is the key for the manipulator to complete the grasping task.However,for the task of grasping an unknown object in an unstructured working environment,the accuracy and real-time performance of the current grasping detection algorithms still need to be improved.In order to solve the above problems,this paper takes the grasping detection algorithm as the research object,proposes the grasping detection algorithm based on anchor mechanism and grasping detection algorithm based on key point estimation respectively,and carries out experiments in the simulation environment to verify the algorithms.First of all,the framework of robot autonomous grasping system was established and each module was introduced.The object perception module based on Kinect V2 was realized.The principle of pinhole imaging,the principle of distance measurement of structural light and the mapping relationship between infrared image and color image of depth camera were studied and the calibration was completed.In the grasping detection module,the grasping prediction representation and Cornell grasping dataset are studied.We also built the corresponding bolt grasping dataset for the stacked bolt grasping detection problem.According to the transformation relationship between pixel coordinate system and robot base coordinate system,the mapping of two-dimensional grasping representation to three-dimensional grasping pose was completed in the grasping planning module,and the kinematics analysis of UR5 manipulator was carried out.Secondly,aiming at the problem that the current grasping detection algorithm was not accurate enough in predicting the grasping orientation,an improved grasping detection algorithm was implemented by introducing oriented anchor to Faster RCNN network.It implemented an anchor-based grasping detection algorithm.In this algorithm,the idea of clustering algorithm was used to set the hyper parameters of oriented anchor,meanwhile,the anchor matching mechanism and the loss function were optimized to improve the performance of the grasping detection algorithm.The algorithm was trained and verified on Cornell grasping dataset,it achieved a 98.1% detection accuracy and a detection speed of 17.13 FPS.Thirdly,aiming at the problem of slow detection speed of the current grasping detection algorithms,this paper implemented a new grasping detection algorithm based on key point estimation by improving the Center Net.In the feature extraction stage of the network,the feature images of each stage were fused by the method of feature fusion to reduce the feature loss.Then a branch of grasp orientation prediction was added on the original network model,and the original loss functions of the network were improved,which made the model suitable for grasping detection tasks.The algorithm was trained and verified on Cornell grasping dataset,it achieved a 97.6% detection accuracy and a detection speed of 42 FPS.Finally,a robot grasping simulation experiment platform based on ROS-Gazebo is built to carry out the grasping simulation experiment.Using Move It! Planning the grasping movements of the robot and completing the hand-eye calibration of the vision robot system.In the simulation grasping experiment for daily necessities,the accuracy of the grasping detection algorithm based on anchor was 88.4%,and accuracy of the grasping detection algorithm based on key point estimation was 87.5%.In the grasping experiment for stacked bolts,the grasping detection algorithm based on key point estimation was used,and the experimental result told that the accuracy of grasping using this algorithm was 84.9%.The robustness and practicability of the robot automatic grasping system were verified by these experiments. |
PDF Full Text Request