| Due to the unique growth environment of Zanthoxylum,it is currently primarily harvested through manual picking methods.To address the challenges related to poor working conditions,worker vulnerability to injuries,and labor shortage during the harvesting process,this study focuses on researching key technologies,such as target recognition and localization,based on the overall design of a Zanthoxylum harvesting robot.Field experiments are conducted to test and analyze the performance of the Zanthoxylum harvesting robot.Based on functional requirements,a complete system for the Zanthoxylum harvesting robot was designed,including the identification and positioning system,end effector,manipulator,and walking system.The key components were designed and parameterized accordingly.The STM32 was chosen as the controller for the end effector,which was designed as an integrated grasping and cutting mechanism,and connected to the industrial computer via serial communication,to achieve the harvesting and picking of Zanthoxylum fruit.A mature Zanthoxylum fruit dataset was constructed by collecting and annotating images of green and red mature Zanthoxylum fruits and performing data augmentation.Four typical object detection networks,Faster R-CNN,SSD,Center Net,and YOLOv5,were trained and tested on the dataset.The average detection accuracy of the four networks from high to low was YOLOv5,Faster R-CNN,SSD,and Center Net.The m AP of the YOLOv5 network was 85.3%,and the detection speed was 97.1 frames/s.To address the issues of the original YOLOv5 network in detecting mature Zanthoxylum fruits,the network was improved by introducing attention mechanisms to enhance the region of interest,using DCNConv2 to introduce offset in the receptive field to extract features of irregularly shaped Zanthoxylum fruits,and optimizing the bounding box loss function based on Wise Io U to improve the overall detection performance.Through heatmap comparison and ablation experiments,the improved YOLOv5 model’s performance was evaluated,and its m AP value was 90.4% with a 5.1% increase over the original model,while the average detection accuracy of the model on the separate test set with occlusion and overlapping Zanthoxylum fruits was improved by 4.7% compared to the original model,with no significant change in detection speed.Based on the improved YOLOv5 detection network output of the mature Zanthoxylum fruit position information,the Zanthoxylum fruit area was cropped as the region of interest for the cutting position.U-Net,Deep Lab V3+ original network,and a lightweight Mobile Net V2 network-based Deep Lab V3+ model were trained.Among them,U-Net has the highest m PA and m Io U,but there is significant overfitting.Mobile Net V2-Deep Lab V3+improved m PA by 0.1% and m Io U by 1.4% compared to the original Deep Lab V3+ network.The model size was reduced to 10.7%,and the segmentation speed was 50.3 frames/s,which is 1.96 times that of the original network.Therefore,Mobile Net V2-Deep Lab V3+ is more suitable for deployment on mobile devices and can accurately locate the cutting position of the Zanthoxylum fruit mother branch by achieving precise image segmentation between the fruit and the mother plant.A field experiment was conducted to test the field passability of the Zanthoxylum(Sichuan pepper)picking robot,the cutting and gripping performance of its end effector,and the generalization ability of its object detection model.The experimental results showed that the Zanthoxylum picking robot can travel normally in field environments with a slope no greater than 27°.The end effector can cut and grip Zanthoxylum mother branches with a diameter of less than 5mm.The visual system can detect Zanthoxylum fruits in real time in field environments and accurately locate the cutting position.The detection speed and segmentation speed on mobile terminal are 89.3 frames/s and 48.6 frames/s,respectively.The results of this study can provide technical support for the design of Zanthoxylum picking robots. |
PDF Full Text Request