Research On Autonomous Pruning UAV System Based On Branch Recognition

With the continuous progress of science and technology,the application of UAV is developing in the direction of multi-function,intelligence and autonomy.The working mechanisms that UAVs can carry,such as vision mechanisms,weapon equipment,picking devices and spraying devices,all illustrate the scalability of UAV applications.Among them,the application of drones whose work task is plant protection has been relatively perfect.Through the advantages of high-altitude operation of drones,manual labor has been gradually replaced,and functions such as spraying liquid medicine,monitoring and early warning can be realized.However,redundant branches,dead branches and disturbing branches also need to be pruned during plant growth,which is basically manual operation,which is costly and dangerous.For this reason,this paper designs a drone that can prune branches in the air.The drone has the functions of branch recognition and autonomous navigation,and can complete the task of aerial branch pruning.This paper first proposes a pruning UAV system scheme,then introduces and designs a deep learning algorithm,and proposes a pruning UAV target detection algorithm based on deep learning.An optimization scheme of target detection algorithm based on Attention-Center Net-L is designed to identify redundant branches that need to be detected.A path planning algorithm based on artificial potential field and dynamic constraints is proposed,and the trajectory of the UAV for tree branch pruning is determined.Finally,the rationality of the design is verified by simulation experiments.The specific work is as follows:(1)According to the application environment of pruning branches,a pruning UAV system is proposed.Firstly,the UAV structure is designed using rhino3 D,and the fourrotor UAV structure is determined,and the portable pruning working mechanism is designed.Then on the basis of the structure,the branch identification module and the operation module of the UAV are designed.The branch identification module adopts a system structure based on the operation of the upper and lower computers,and is used to identify redundant branches in the working scene.The operation module mainly uses the pruning working mechanism to realize the function of pruning,and complete the pruning task planning of the working system when it is close to the target branch.The ultrasonic sensor and the electric pruning mechanism are used on the module to coordinate the work.(2)A target detection algorithm model based on deep learning is constructed,and a deep learning-based pruning UAV target detection algorithm is proposed that is suitable for the design goals of this paper.On the basis of the target detection algorithm based on traditional neural network,the Center Net target detection algorithm scheme with relatively excellent performance is given.The residual network module Res Net50 is integrated into the backbone extraction network of Center Net,and the network model of Center Net-Res Net50 is built.Through the process analysis of model training and prediction,the decoding and reasoning of the network model are realized.(3)Optimize the deployment of the branch detection model based on Center Net,and propose an optimization scheme for the target detection algorithm based on AttentionCenter Net-L.Firstly,based on the Center Net center point target detection algorithm,in order to meet the needs of the detection target,the Res Net50 backbone extraction network in the target detection algorithm is optimized.Put the multi-introduced batch normalization(BN)and activation function Re LU into the input of each residual module of the backbone extraction network,and move the activation function Re LU of some residual output up to avoid the feature transfer process.The scattered data constitutes the Res Net50-L optimization scheme,which accelerates the training speed of the network,optimizes the convergence of the deep network in the process of processing data,and prevents the occurrence of gradient explosion and disappearance.Based on the improvement of the backbone extraction network,we chose to add the CBAM attention mechanism module in the process before upsampling,focusing on the accurate identification of the target class and preventing errors in the classification situation,which constituted the Attention-Center Net optimization scheme.(4)The path planning and design of the pruning UAV is combined with the use scene,and a path planning algorithm based on artificial potential field and dynamic constraints is proposed.Based on the dynamic modeling of the UAV body,the PID algorithm is used to realize the mathematical modeling of the UAV’s underlying flight controller,so that the UAV remains stable in the flight path based on the trajectory planning.Analyzing the needs of the working scene,through the collection of tree data,the virtual model and obstacle model of the tree are constructed,the scope of non-pruning is standardized,and the path planning scheme design of the UAV based on the real scene is completed,and the local area is determined.Path following strategy.Through the action of the mutual repulsive force under the artificial potential field algorithm,on the basis of obstacle avoidance,the dynamic constraint of the virtual centripetal force is added,so that the trajectory of the UAV will not deviate.(5)Integrate the two algorithm optimization schemes of Res Net50-L and AttentionCenter Net to form the optimization scheme of the Attention-Center Net-L target detection algorithm,and complete the actual embedding of the above optimization scheme through the Python language and the operating environment of Py Torch.And the VOC data set is used to verify the model.The results show that the accuracy of target recognition is improved under the premise that the training time is almost unchanged,and the practicability of the optimized target detection effect is verified.On this basis,according to the design task of this paper,a sufficient amount of tree image data is collected in the environment of the school,and the recognition experiment of the pruned part of the tree is carried out.Make a data set,and finally train the data set,build the model data in the network,and finally complete the identification of pruned branches.According to the design parameters of the pruned UAV,the UAV model was built in Solidworks,imported into Simulink to build the pruned UAV controller module,and the flight control of the UAV operation was simulated and verified.According to the above design parameters,the simulation model of UAV and obstacles is built on the Matlab simulation experiment platform,and the simulation experiment research on the path following strategy of pruning UAV is completed,and the path based on artificial potential field and dynamic constraints is verified.Feasibility of planning methods.From the functional point of view,the autonomous pruning drone based on branch identification proposed in this paper can be applied to aerial operations such as garden pruning,plant protection pruning,and high-altitude interference branch pruning,and has broad application prospects.