Research And Design Of Vision System Of Fruit Picking Robot

China is a major agricultural producer in the world,and the picking industry is an important link in agricultural production,which requires a large number of manpower to quickly and intensively pick fruits in the designated season every year.With the sharp decline in the number of agricultural workers and the rise in labor costs in China,it is urgent to carry out the research on picking robots to achieve intelligent picking,in which the accuracy of fruit target location detection and classification identification is a crucial link.Therefore,this thesis studies and designs the vision system of fruit picking robot in the natural environment,laying a foundation for the realization of high-performance picking robot.The main research work and research results obtained in this thesis are as follows:(1)According to the demand of fruit picking robot vision system in agricultural picking,the fruit image database was built,and the fruit target location and recognition were completed.First,the Faster R-CNN algorithm is studied for the localization and detection of fruit targets.Then,VGGNet algorithm of Convolutional Neural Network has many network layers,large amount of computation and long operation time,according to the actual scenario requirements in this thesis,an improved VGGNet algorithm is proposed,using Adam optimization algorithm to replace the traditional stochastic gradient descent algorithm,used to update the network weight parameters and independent learning rate adjustment,finally realizes the classification recognition of the target of the fruit.(2)The hardware platform and software environment of the visual system are built,and the indoor simulation experiment is carried out.The visual system program was written under the framework of Keras,and the image data of 5 kinds of fruits were trained and the model was generated.After successfully testing the visual system on PC,the program was transplanted to the Raspberry PI 3b+ development board for simulation experiment based on embedded platform.The experimental results showed that the accuracy of fruit localization detection and classification recognition could reach 94.77%.(3)Based on the completion of the visual system,the system was applied to the 6-DOF picking manipulator arm for indoor picking experiments.The experimental results show that the accuracy of cherry tomatoes location detection and classification recognition is 97.5%,and the grasping success rate is 92.5%.In conclusion,the visual system designed in this thesis can be applied to the location detection and classification recognition of various fruit images in the natural environment,and it is effective and practical in the picking process,providing strong support for the research and development of fruit picking robots.