Study On Visual Recognition And Picking Technology Of Tender Tea Buds

With the continuous improvement of people’s quality and drinking tea demand,the demand for tea tender buds is getting bigger and bigger.The research work of the intelligent picking of tender tea buds is imminent,and this paper is based on target identification technology and robot technology to develop tender tea buds visual identification and picking technology,and developed tea buds picking robot model.During the study,in order to make the equipment can accurately pick the tea buds that meet the requirements,we focus on the visual recognition of the tender buds,coordinate output,path control of the manipulator and picking sequence.This paper focuses on the speed and accuracy the identification of tea buds picking point,coordinate sequence control of manipulator motion and mutual communication feedback.First of all,according to the images taken in the tea garden and the experience of picking buds,aiming at the growth characteristics of tea buds and the hand eye cooperation relationship when picking buds,this paper proposes a "two-step" picking point recognition and picking scheme based on the combination of machine learning and deep learning.In the first step,the global camera installed on the tea picking robot is used to capture the whole image of tea,and the machine learning method is used to extract the tea bud foreground,so as to find the initial positions of all the buds in an image,and the manipulator reaches the vicinity of a single bud one by one according to the position information.In the second step,the local camera located on the manipulator captures a single tender bud,and the deep learning method is used to identify the specific picking point of a single bud and output coordinates.Further,in order to shorten the tea picking time,this paper attempts to remove the process of image segmentation of the buds foreground and directly identify the tea picking point by "one step" method.Therefore,the T-yolo algorithm model is proposed for the identification of the picking points of tea buds.The original image of tea is directly used as the input,mosaic enhancement and other operations are added to the data set,the detection speed is improved by combining with the cspdarknet-53 of deformable convolution,and the shape of tea buds is also better adapted.The combination of SPP and PAN is used to increase the receptive field and aggregate information.Finally,the output is aggregated into a scale by using hypercolumn composition method.It is proved that the performance of the T-yolo model can meet the requirements of tea buds recognition,and it has a good performance in detection time and accuracy.How to unify the coordinate system and communication with robot needs further research.Finally,the paper combines the visual recognition module of bud with machine learning and deep learning,the mechanical structure module with delta parallel mechanism as the main body,the path planning module of robot with improved ant colony algorithm and the communication module of setting up the the transmission of buds coordinates and road control instruction by Socket to complete the overall design of tea picking robot.The prototype model is made and the indoor tender tea picking experiment is carried out.