Single Bud Detection And Accurate Picking Technology Of High-quality Tea Based On Machine Vision

Data from China Tea Circulation Association indicates that the total area of tea plantations nationwide will be about 3.2 million hectares by the end of 2022.With the increasing shortage of labor in tea plantation operation and the scale of tea plantation due to the centralization of land management in the future,the application of advanced famous tea picking robots to replace the backward manual picking or undifferentiated mechanical harvesting operation has become an inevitable trend for the harvesting of famous tea in tea plantations.The existing famous tea picking robots are limited by the low efficiency of online detection of tea buds and poor accuracy of picking point positioning,and there is less research on tea bud picking strategy and energy consumption optimization.Therefore,how to develop an accurate and efficient vision system to detect tea buds in tea fields and locate picking points,and at the same time pick tea buds in a better sequence with high efficiency and energy saving is an urgent problem to be solved.The main research of this paper is as follows:For the problem of low efficiency and poor accuracy of tea bud detection,the YOLOv3 network model is improved and a fast detection method for single buds of famous tea is proposed.Pictures of tea buds are collected in tea gardens to establish a tea bud detection dataset.The classical YOLOv3 network model was improved by adding a pyramid pooling module,improving the multi-scale fusion detection module,and optimizing the anchor frame and loss function.The improved network achieves 81.42%detection accuracy and 15.87 frames/sec,respectively,with stronger small target detection capability and faster detection speed,which lays a good foundation for tea bud picking area identification.To address the problem of poor accuracy of tea bud picking point localization,a threedimensional localization model of tea bud picking point based on PSPNet network and depth camera is proposed.A tea bud picking area recognition dataset is established,and the PSPNet network-based tea bud picking area recognition model is built and trained.The average accuracy of one bud picking area recognition is 80.76%and the average cross-merge ratio is 72.60%.The accuracy and reliability of picking point recognition and localization were proved by using depth camera combined with center of mass method to find the coordinates of picking points in the camera space and setting up tea bud picking point localization test.For the problem of how to pick tea buds efficiently,a parallelogram picking strategy and tea bud picking sequence planning method are proposed.Based on the objectives of maximum efficiency and minimum energy consumption,a path-energy consumption integrated optimal objective function is designed.The classical ant colony algorithm is improved,and the improved ant colony algorithm has a better finding ability than the classical ant colony algorithm in both the cases of moderate amount of tea buds and large amount of tea buds.In the simulation experiments with a moderate amount of tea buds,the tea bud picking sequence planning method in this paper reduced the energy consumption by 19.65%compared with the shortest path planning method,while the total path length increased by 3.01%,and achieved the maximum range of the picking robot while working efficiently.In order to verify the feasibility of the research method,a prototype tea picking robot was piloted and an intelligent picking control system for the tea picking robot was also developed.In the picking sequence planning test,the picking sequence planning method in this paper increased the picking time by 0.28%and reduced the energy consumption by 12.5%compared with the shortest path planning method in the case of moderate amount of tea buds.In the tea field picking test,the success rate of tea bud detection was 90.02%,and the average time consumed for each tea bud picture detection was 0.0627s;the success rate of tea bud picking point localization was 84.21%,and the average time consumed for each tea bud picking point localization was 0.0158s;the success rate of tea bud picking was 77.30%,and the average time consumed for each tea bud picking was 2.01s,which can satisfy the operation demand of tea picking robot in tea field.