Research On Fast Picking Technology Of Apple Harvesting Robot

In China agricultural production,as one of the important leading industries,apple annual output and export value are ranked first in the world.In the whole production process of apple,the labor cost of picking is high and the cycle is short.At the same time,the decrease of agricultural population leads to the shortage of labor force engaged in apple production,which makes people's demand for automatic machinery more and more urgent.Automatic Apple harvester can replace most of the work in the process of apple picking,and realize the automation,mechanization and centralized management of apple picking.Supported by the project of the national natural science foundation of China(31571571)"research on efficient machine picking method of apple in a multiillumination environment based on fast visual servo control",this paper researches on apple picking technology based on deep learning and visual servo control in apple harvesting robot.This paper takes all kinds of apples in the complex environment of the orchard as the research object,the research focus is the rapid identification method of the target apples and the quick picking method using the pneumatic device.Respectively based on the research of the vision algorithm increases the speed of recognition,based on the research of the servo control system improve the crawl speed,a combination of common improve the picking speed of the robot.The main work completed in this study includes the following points:1.Selection of visual algorithm in apple harvest robot.Due to the unstable characteristics of the target apple in the complex environment,the traditional visual recognition method based on color,shape,and texture features cannot fully identify all the apples in the environment.Therefore,a method based on deep learning is proposed to identify the apple position.In this study,fruit images collected from apple orchards are used to make data sets.Meanwhile,the deep learning framework Darknet is used to build the deep neural network YOLO(You Only Look Once),to adaptively learn the characteristics of target apples and realize the identification of apples in complex environments.In order to achieve the purpose of fast recognition(meet the real-time requirements),the scale of neural network is reduced without affecting the accuracy of the algorithm,so that the algorithm can detect apples in real-time.2.Creation of deep neural network structure and adjustment of super parameters in the vision system of apple harvest robot.In order to train a deep neural network that can correctly recognize the position of apples,good data sets,hyperparameters,and neural network models are needed.All the samples used in the data set are from the actual orchard rather than the laboratory.Such samples with large difference in feature distribution can train a neural network model with higher robustness,and avoid the model's overfitting of the training set samples,which makes it unable to detect the samples outside the data set.Hyperparameters involve learning rate,excitation function,sample input size,etc.Good hyperparameters can make the gradient in the training process rapidly decline,and the training curve will not oscillate or stop at a saddle point.In the process of model training,different hyperparameters are used to conduct training experiments by controlling variables,and the appropriate hyperparameters are selected by referring to the changing process of model loss.The size of the neural network model will directly affect the running speed of the overall algorithm,experimental classes are used respectively to VGG and Resnet network training,training for each of the network model,the size of each are not identical,through the different specifications network on the test set of the mAP(mean Average Precision),Accuracy,Precision and Recall,assessment of IOU(Intersection Over Union),testing time.Finally,the appropriate neural network structure is selected and the trained model is used in the algorithm of the visual recognition system.3.Optimization of apple grasping efficiency in the picking process of apple harvesting robot.The research of harvesting robot adopts pneumatic device to optimize crawling speed.Devices involved in the picking process include Festo pneumatic push rod,photoelectric gate sensor,cutter and end-effector.Festo pneumatic push-rod is used to propel the end-effector during the grasping process.The photoelectric gate senses the apple's trigger signal and activates the end-effector ground grabbers.After catching the apples,the then cutter separates them.4.Evaluation of picking speed in an Apple Harvest-robot quick Grab experiment.This paper evaluates the apple continuous speed and precision of the grab,verify the vision algorithm and the optimization of pneumatic control flow can significantly improve the whole process of picking fruit.In the automatic apple picking experiment,the average apple picking time is 7.81 seconds,in which the visual algorithm recognized the apple for 81 ms.The optimized apple grab time of pneumatic system is 2.70 seconds and apple recovery time is 0.64 seconds,which is significantly improved compared with the previous experiment.