Citrus Picking Path Planning And System Implementation Based On 3D Point Cloud Model

Citrus are widely planted in China.At present,the picking of citrus fruits is still mainly by manual picking.With the increasing cost of labor,there is a shortage of labor for citrus picking operations in the future.Using mechanized picking technology instead of manual picking can promote the upgrading of industrial structure and solve the labor shortage in citrus picking operation.Combined with the development of intelligent agriculture and the construction of unmanned orchards,this paper designed a citrus picking platform based on a dual robot system,and studied the problems of the dual robot system calibration,citrus recognition and positioning,and citrus picking path planning in the process of citrus picking.The main contents are as follows:(1)A citrus picking platform with dual-robot system was designed,which is composed of citrus model tree,vision system and picking robot system.The 3D modeling of citrus tree is completed by a 3D scanner installed at the end of the scanning robot,and the picking of citrus is completed by a picking manipulator installed at the end of the picking robot.(2)The kinematics of the robot is analyzed,the kinematics equation of the robot is established,and the inverse kinematics solution of the robot joint Angle is completed.The calibration problem of dual-robot system is transformed into a mathematical model for solving unknown matrices X,Y and Z in matrix equation AXB=YCZ.The closed form method and iterative method are used to solve the translation vector and rotation matrix of unknown matrices X,Y and Z,and the calibration experiment of the dual-robot system is designed and completed.(3)The coordinate transformation process of point cloud generated by 3D scanner is completed.A 3D model of citrus tree is established in the base coordinate system of picking robot,and the point cloud model is pretreated with down-sampling and de-noising.The number of point cloud models after down-sampling is about 20%of that before processing,and the de-noising of point cloud model made a large part of outliers in the model removed.3000 images containing citrus are trained by YOLOv5 algorithm model.The Precision value of trained YOLOv5 model on the test set is 94.4%,Recall value is 99.0%,AP value is 99.2%,and F1 value is 98.0%.The training model is used to complete the recognition of target citrus in the image.The pixel coordinate system of the image is established and normalized.According to the matrix relationship between the normalized coordinate system and the base coordinate system of the picking robot,the point cloud cluster of the target citrus is further determined,and the mean coordinate of the point cloud cluster is calculated.The mean coordinate is used to represent the spatial position of the citrus,and the location of the target citrus is completed.(4)Based on the principle of shortest path,the citrus picking path planning problem is transformed into a mathematical model of single traveling salesman problem,and the traveling salesman problem is solved by genetic algorithm.The optimal citrus picking sequence is obtained by iterative calculation.Ten groups of citrus picking experiments are designed and completed,each group contained 10 citrus.In the picking experiment,the total picking time averaged 158.9 seconds,and the picking success rate is 82%.This paper designed the citrus harvest scheme based on dual robot system,completed the double robot calibration system,adopt the way of deep learning and point cloud processing,respectively,to identify and locate the target citrus,planning the citrus picking order and use the completed on citrus picking,picking manipulator for unmanned orchard picking technology to provide theoretical basis and technical guidance.