Reserch On Object Recognition And Pose Estimation Technology Based On The Robot Eye-in-Hand System

Recent years have witnessed great development of E-commerce,and the demand for intelligent goods sorting in the warehouse logistics field increases day by day.Combination between robot manipulation and vision algorithms can greatly improve the intelligence of the robot system.This thesis is concerned with the perception algorithm research for object recognition and pose estimation based on the robot eye-in-hand system.The thesis can be mainly divided into the following parts:(1)Based on the fully convolutional neural network,we perform image segmentation on the RGB image of a RGB-D camera and realize pixel-level classification for each kind of object.During the training process,we use frame-subtraction method to simplify the image sampling and reduce workload of manual labelling.Contrast experiment shows that FCN enjoys more advantages in terms of time efficiency and scalability.(2)After semantic segmentation for the color image in a RGB-D camera,we utilize point cloud registration to perform pose estimation for each objects.Point cloud registration adopts the coarse-to-fine scheme.The coarse registration is realized by principle component analysis(PCA)or point pair feature algorithm(PPF),while the fine registration is realized with the help of iterative closest point algorithm(ICP).We introduce principles and implementation details of PCA and analyze the advantages,disadvantages and strategy for applying PPF in real scenarios.Finally,to achieve fast and convenient point cloud modelling of the target object,we design a framework based on the popular visual odometry or visual SLAM to meet the modelling demand.(3)In order to take full advantage of the eye-in-hand system and increase accuracy for object recognition and pose estimation,this thesis discusses the multiple view pose fusion problem.Pose fusion is realized through pose regression.Pose regression can be conducted on the matrix F norm sense and the lie algebra sense.We introduce the related math background,derive formulas for both methods and develop a strategy to conduct multiple view pose estimation on the robot eye-in-hand system.Numerical experiment shows that pose fusion on the matrix F norm has the best performance in terms of efficiency and accuracy,which makes it suitable in real applications.Experiment on the robot platform demonstrates that the multiple view pose estimation strategy proposed by this thesis can effectively increase the accuracy for pose estimation.(4)To combine image segmentation,pose estimation and robot control for the intelligent object picking task,we design the software framework based on the robot operating system(ROS).Nodes are designed for visual algorithms and robot control and function coupling between nodes is removed as much as possible to increase readability and maintainability of the program.Experiment shows that the software framework proposed in this thesis can fulfill demands of the intelligent picking task of the robot eye-in-hand system.