Visual Recognition Positioning And Motion Planning Of Intelligent Mobile Manipulator

In the context of a large agricultural labor gap and increasing picking costs,the use of picking robots can significantly increase efficiency and reduce costs,and realize automation and intelligence of agricultural production,which has important practical significance and broad application prospects.The picking robot often adopts the technical solution of the mobile chassis and multi-degree-of-freedom manipulator.This thesis studies the key technologies of the picking mobile manipulator from the three aspects of vision system design,apple recognition and positioning,and manipulator motion planning.First,the design of the robot vision system is completed.The depth camera is installed at the end of the robotic arm using the "eye-in-hand" method,and camera calibration and hand-eye calibration are carried out.An improved internal parameter optimization method of an improved quantum genetic algorithm is proposed to improve the accuracy of camera calibration.Secondly,a lightweight convolutional neural network design and intelligent target recognition method are proposed.Improve the design of YOLO-V3 network,design a feature extraction network structure with homogeneous residual blocks in series,simplify the feature map scale of target detection,and replace ordinary convolution with depthwise separable convolution.This thesis proposes a multiobjective loss function that combines the mean square error loss and the cross-entropy loss.Develop a crawler program to obtain training data from the Internet and label it,use data augmentation technology and normalize the data.A multi-stage network training method based on SGD and Adam optimization algorithms is proposed.Calculate the center position of the apple using a least squares fit.Thirdly,the research of manipulator motion planning is carried out.Based on the DH parameter method,the kinematics model of the UR5 manipulator is established,and forward kinematics and inverse kinematics are analyzed.Kinematics simulation and workspace simulation are completed.Aiming at the scene of automatic picking after fruit recognition,an improved multi-stage,collision-free manipulator motion planning method for RRT-Connect is proposed,which realizes the manipulator’s motion planning in the joint space and performs manipulator collision detection in the workspace.Finally,in order to verify the effectiveness of the above research methods,a manipulator visual recognition control system is developed on the ROS framework,and apple recognition experiments under complex scenarios is performed on workstations and embedded devices.A prototype of a picking robot is built using a mobile chassis and a UR5 robot arm.Apple recognition,grabbing,and placement experiments are performed in a laboratory environment.The experimental results show that the detection speed and accuracy of the apple detection method are significantly improved,and the manipulator planning method can perform efficient planning in an obstacle environment,which will lay a good technical foundation for the field application of apple picking robot embedded computing resources.