Design And Experimental Research Of Tree Fruit Picking Equipment Based On Machine Vision

To rejuvenate the nation,rural areas must be revitalized.The state has repeatedly emphasized the priority of implementing the rural revitalization strategy.Fruit planting industry occupies an important position in agricultural economy.At present,most domestic fruit picking relies on manual labor,and the labor cost in the whole fruit industry accounts for a large proportion.With the acceleration of population aging,the labor cost increases,and the use of intelligent technology and equipment with high efficiency and low cost to realize the automation of agricultural production is an inevitable trend of future development.In this paper,according to the domestic orchard planting environment and industry conditions,research and design of a tree fruit picking robot based on machine vision,which consists of 4 degrees of freedom robot arm,end holder,RGB-D vision system,can provide key technical support for the automatic harvesting work of several main tree fruits,to provide reference for the realization of intelligent picking robot early commercialization.The main research contents are as follows:(1)Analyze the domestic orchard planting environment,fruit growth state and geometric physical characteristics,study the overall design scheme of two kinds of picking robots,and carry out comparative analysis,complete the design and selection of the main mechanism in the execution system,including the selection of walking mode,the establishment of the form and freedom of mechanical arm mechanism,the choice of two end-effector schemes.At the same time,the selection of industrial camera in the vision system is carried out.The distortion calibration of RGB-D camera and the fusion of depth image and color image are tested and analyzed.The structure composition and working principle of the corresponding control system are designed according to the functional requirements of the robot,and the power system is analyzed.(2)By analyzing the features of apple fruit image,a series of methods such as conversion from RGB color space to HSV color space,separation of HSV three-channel,K-means clustering segmentation,threshold segmentation,morphology open operation and contour extraction are used to segment and process the apple image.The correct recognition of apple targets in independent,adjacent and overlapping states is realized,and the average recognition rate can reach 93.72%.(3)According to the hardware and computing power limitations of the host computer of the picking robot,the YOLOv3 network model is lightened,and the collected apple data set is expanded by data enhancement technology.The trained YOLOv3 network model can effectively identify the Apple target,with the accuracy of 97.24%,the recall rate of 92.01%,the F1 value of 94.63% in the test set,and the detection time of a single image of about0.13 s.The comprehensive recognition ability of the network model is better than the traditional image segmentation algorithm.It can fully meet the requirements of automatic apple picking and identification.(4)The robot picking test platform was built,and the overall performance of the picking robot was verified by picking in the real environment.The results showed that the average single recognition time of the picking robot was 1.77 s,the average single picking time was 10.14 s,and the average success rate of picking apples was 90%,which basically met the demand of automatic fruit picking.The feasibility and performance of the proposed algorithm are further verified.