| Kiwifruit is more difficult than other species on pollination,since not only it’s dioecious plant without synchrony of flowering time,but also its flowers lacking nectar are not attractive enough to pollinators.Artificial pollination overcame difficulty and became a critical technique to increase kiwifruit quality.However,there is no doubt that artificial pollination is a labor-intensive process resulting in high labor costs and time-consuming.Besides,fruit farmers and casual workers lacked agronomic knowledge training,and their artificial operation of pollination was highly random,which affected kiwifruit quality.A few reported pollination robots were not equipped with intelligent control system and were combined just a little knowledge of agronomic pollination,where there were problems such as poor pollination quality and waste of large amount of pollen.Based on in-depth study of kiwifruit pollination agronomy,this paper combined female flowers growth stages and states and their distribution information to develop accurate pollination device of kiwifruit flower based on visual perception and air-liquid spray,which was hopeful to achieve precise pollination of dominant female flowers in the canopy and improve pollination quality.The main research contents and conclusions are as follows:(1)Overall scheme design of accurate pollination device.According to kiwifruit planting characteristics and canopy female flower distribution and agronomic characteristics,in the lower canopy area,kiwifruit female flower information was obtained,and then the pollination operation was carried out for dominant female flower.Based on comparative analysis of advantages and disadvantages of different pollination methods,air-liquid spray pollination method was selected because it was highly oriented and easy to control pollen volume.Overall scheme and workflow of the pollination equipment were determined based on the workspace information and pollination operation requirements under trellis cultivation mode.(2)Kiwifruit flower canopy images collection and processing.Based on kiwifruit cultivation patterns and growth characteristics,image acquisition method,equipment and time were determined.According to kiwifruit pollination agronomy and the distribution structure of kiwifruit canopy branches and flowers,classes of annotation for the final floral canopy image dataset were determined,which were based on the characteristics of female flowers at different growth stages and the change of state after female flower pollination.Then,raining and validation datasets were randomly divided into 4:1 ratio according to the needs of target detection,and dataset augmentation method of floral canopy dataset was determined.Images were annotated as dataset for deep learning network,where the training dataset and validation dataset were randomly divided according to a 4:1 ratio,and the augmentation method for the floral canopy dataset was determined.All default format annotation files were converted to proprietary format annotation files for YOLO network,and the training and validation datasets for YOLO network were created for subsequent network training and evaluation.(3)Construction of visual perception system for kiwifruit flower canopy based on YOLOv5 l.YOLOv5l was selected for training and evaluation of the kiwifruit canopy multiobjective dataset based on the experimental hardware platform conditions,model size,detection accuracy and portability.APs of all 14 classes of the network model trained by YOLOv5 l were above 95%,and its m AP was 98.4% with excellent detection performance.Moreover,its model size was only 89.4 MB,and its average detection speed was 15.50 ms per image with 4068 × 3456 pixels under the training platform.According to canopy agronomic characteristics of the standardized kiwifruit orchard trellis-type planting pattern,decision method of dominant female flower based on growth stages and location distribution of female flowers on fruiting branches is proposed to achieve accurate decision method of dominant female flowers on fruiting branches.In order to avoid the influence of camera lens parameters on positioning accuracy,the Real Sense D435 camera was calibrated using the Zhang Zhengyou’s camera calibration algorithm to obtain its camera parameters.Interconversion of flower coordinates from pixel coordinate system to camera coordinate system was achieved based on the RGB-D camera spatial coordinate conversion method.Conversion method from camera coordinate system to world coordinate system was determined.(4)Development and experiment of air-liquid spray pollination device.According to the agronomic and pollination control requirements of kiwifruit cultivation,the plan of air-liquid spray pollination device was determined,because of its the advantages of easy control of pollen spray volume,strong guidance,good uniformity of atomization and uniform droplet size.The design and hardware selection of the control system were completed,whose circuit schematic diagram was drawn.The experiment platform of air-liquid spray pollination characteristics was constructed,and the three-factor five-level quadratic orthogonal rotational combination experiment was designed by selecting spray air pressure,flow rate and spray distance as variables.The optimal spray parameters of air-liquid spray nozzle were determined by using the “Design-Expert” data processing tool: spray distance of about 25 cm,pollen flow rate of about 44 m L/min,and atomization air pressure of about 58 k Pa.(5)Constructing precision-to-target spray pollination equipment and field pollination experiment.Specific design of target pollination equipment,construction of whole machine and design of control system were completed,and then pre-experiment of target spray and field pollination experiment were carried out to verify the effectiveness of the pollination equipment.The camera external parameters were determined according to the relative position of the camera and the first axis of the robotic arm,and then positioning errors were measured,which met the pollination positioning requirements of this paper.An indoor pre-experiment of target spray was completed,which obtained elapsed time of target spray operation.The visual perception module taking about 1 s,while the average pollination time for a single flower was about 2 s.In the field orchard pollination experiment,the fruit set rate was 88.5% in the area pollinated by air-liquid target spray method,which was only 4% lower than that in the area pollinated by hand-dispensing and 6.7% higher than that in the area pollinated by hand-spray.The average fruit weight in the area pollinated by the selected dominant female flowers based on air-liquid target spray was only 1.7 g less than that in the area pollinated by hand-dispensing,and the average fruit diameter was close to that in the area pollinated by hand-dispensing.The air-liquid target spray pollination method could save 39% and 42% pollen compared with hand-dispensing pollination and hand-held sprayer pollination,respectively.In summary,this paper developed accurate pollination device of kiwifruit based on vision perception and air-liquid spray for standardized kiwifruit orchard.The visual perception module based on YOLO5 l achieved precise detection of canopy images,decision method of dominant female flower and 3D positioning.The spray module based on air-liquid nozzle provided a reliable pollination component for the whole machine.The robotic arm targeting module moved the spray module to the designated position and aimed it at the female flowers,thus completing the spray pollination operation.This paper integrated pollination agronomy,computer vision,air-liquid spray nozzle and robotics to perform differential flower targeting and precise pollination of female flowers in the canopy,which contributed to establishing a new model of pollination operation in kiwifruit smart orchards. |
