Research On Pose Estimation Based On RGB Image Manipulator Grasping

With the rapid development of science and technology and the increase in labor costs,the use of robots to replace manual labor has become more and more widespread.At present,most manipulators only perform point-topoint operations mechanically,and can only work effectively when the environment is fixed and the tasks are fixed and repeated.Once the working environment,target status,and grabbing tasks have changed,they need to be reconfigured,which is a lack of flexibility.Therefore,for the manipulators,it is a very challenging problem to correctly identify and effectively grasp the target object with an unknown pose.Thanks to the substantial increase in computer computing power,deep learning technology has achieved rapid development.To improve the autonomy and intelligence of robot operation,the study of combining deep learning technology with robot operation has good academic value and practical significance.In this thesis,the manipulators grabbing based on RGB images is researched in complex scenes.Aiming at the problem that the partial solution obtained by inverse kinematics makes the manipulator not suitable for practical application due to the multi-solution situation of inverse kinematics,this thesis uses a path planning constraint optimization method that combines the joint constraints,workspace constraints and collision detection.This method solves the problem of collisions with the experimental platform during the movement of the robotic arm,and at the same time addresses the problems of complex and changeable motion planning trajectories,redundant trajectories,and excessive motion range.Aiming at the problem of partial occlusion of the object in the process of grasping,this thesis designs a depth convolution neural network based on multi-scale feature fusion to extract the projection feature points of 3D objects.Then proposes a method of using different PNP algorithm according to different projection feature points.In this thesis,Res2 net network is used as the basic framework to extract multi-scale information.The multi-level pyramid pooling is added to obtain context information of different regions.For the problem that the number of projection feature points of the threedimensional object predicted by the model is different,P3 P algorithm is used for the first situation that 4?n?5.For the second situation that n?6,a least square formula is constructed to solve it by taking the solution obtained from P3 P algorithm as the initial value in the Levenberg Marquardt based iterative method.Aiming at the computational complexity and time-consuming problems caused by the accurate model in pose estimation,a simple geometric space body,namely,the bounding box,is used to approximately replacing the accurate three-dimensional target object model.This method simplifies the object model,reduces the amount of computation.Therefore,it has good real-time performance.At the same time,a large amount of synthetic data is automatically generated by the computer simulation,which avoids the timeconsuming and laborious problem of manual generating the data set.The experiments verified that the network trained with synthetic data can also work effectively in real scenes.Finally,this article builds a robotic arm grasping platform based on ROS.The trained model is deployed on the platform for grasping experiments to verify the performance of the pose estimation algorithm in application.Experiments show that the proposed pose estimation method using only RGB image information is competent to estimate the pose of the target object and capture these objects in the actual scene.