Design And Experiment Of Kiwifruit Precision Target Pollination Robot Based On Deep Learning

As the planting area of kiwifruit in China is increasing year by year,the corresponding pollination demand is also increasing.In order to solve the problems of extensive pollination,low pollen utilization,high labor intensity,easy damage to plants,and poor pollination targeting in the current pollination process,this paper designs a kiwifruit precision target pollination robot based on deep learning.Based on the needs of flower recognition and accurate position judgment,this paper optimizes the visual recognition algorithm based on YOLO v5,and combines the double-flow nozzle to carry out multiple sets of parameter comparison tests.Finally,the best spray pollination parameters are determined,and the pollination robot is tested and verified.The main research contents of this paper are as follows :(1)Research on efficient recognition technology of kiwifruit flowers based on improved YOLO v5.In order to realize the accurate recognition of flower overlap and flower tilt angle in natural environment,a kiwifruit flower detection algorithm based on improved YOLO v5 is proposed.In the backbone network,K-means++ algorithm is used to cluster more accurate anchor frames,and CBAM attention mechanism is added to improve its feature extraction ability.At the same time,CIOU is used as a loss function to obtain better regression results.The flower overlap judgment module and the flower tilt angle calculation module are added to the detect function.The experimental results show that compared with Faster-RCNN,SSD,VGG and other algorithms,the YOLO v5 s model has the best recognition effect.The recognition accuracy of flowers and stamens reaches 96.7 % and 91.1 % respectively,and the recall value is high,which can meet the recognition requirements of kiwifruit flowers and stamens in the case of overlap and tilt.(2)Study on dual-flow precision target pollination spray technology.A precision target pollination method based on angle control is proposed.The test scheme of precision target related parameters and the test method of deposition amount and droplet size are designed.The influence of inclination angle on the area of droplet deposition area and the escape rate of droplets is analyzed.Finally,the droplet deposition amount and pollination success rate of precision target pollination method are analyzed through experiments.The results showed that when the pollination hydraulic pressure was 0.070 MPa,the pollination pressure was 0.150 MPa,the pollination distance was 40 cm,and the pollination base time was 3.50 s,the droplet deposition could meet the requirements of full pollination without tilt angle.With the increase of the inclination angle of flowers and nozzles,the maximum droplet escape rate is 27 %.The average droplet deposition in the stamen area of the precision target pollination method at different tilt angles was stable at 43.0～46.5 mg.The average final pollination time required for a single flower was 3.67 s,and the pollination success rate reached 96.7 %.(3)Structure design of precision target pollination robot.According to the working environment and operation requirements of the pollination robot,the modular design of its structure is carried out.The mechanical size is determined according to the installation distance of the camera,and the main driving source is selected.(4)Pollination robot control system design.According to the control requirements of pollination robot,the PCB circuit board for pollination robot is designed and the related components are selected.The robot control system based on STM32 microcontroller is designed to realize the real-time communication between the lower computer and the upper computer,and the visual interface of the upper computer is designed according to the real-time observation requirements of pollination.(5)Static identification accuracy test,mechanical device motion accuracy test and continuous spray pollination test.In order to verify the accuracy of the pollination robot recognition system and the precision of the pollination operation,the above three experiments were carried out.The experimental results show that the accuracy of flower number recognition of pollination robot under different illumination is above 90 %,and the average recognition time is about 180 ms.The average deviation of coordinate search is 4.18 mm,the average deviation of angle search is 2.24°,and the average deviation of pollination distance is 1.59 mm,which meets the requirements of pollination operation.The average values of droplet deposition under the conditions of single flower,2 overlapping flowers,3 overlapping flowers and 4 or more overlapping flowers were 43.4 mg,46.7 mg,49.3 mg and 53.1 mg,respectively,showing a steady upward trend.In the case of random overlap,the average droplet deposition under three light conditions was 45.7 mg,45.9 mg,46.2 mg,the average pollination time of flowers was 8.19 s,8.37 s,8.42 s,and the pollination success rate was 92 %,91 %,93 %,respectively.The pollination accuracy and success rate under different light conditions can remain stable.