Research On Navigation And Inflorescence Identification Method Of Trellised Pear Orchard Incorporating 3DSLAM Technology

In 2021,China’s fruit planting area reached 192 million mu,with a total output of nearly 300 million tons,with the orchard planting area and fruit production ranking first in the world.Fruit cultivation can obtain several times the economic benefits of field crops,has become an important pillar industry for farmers to increase income.However,the comprehensive production efficiency of China’s orchards is low,especially the amount of labor,labor-intensive,tight farming time of fruit and flower management,which consumes a lot of labor.2023 Central Document No.1 and the "Fourteenth Five-Year Plan" to promote agricultural and rural modernization have put forward the requirements of strengthening agricultural science and technology and equipment support,orchard mechanization,intelligent Technology research and development and application,is the fruit industry sustainable development of science and technology guarantee.Therefore,this study focuses on the autonomous inflorescence recognition and accurate positioning required for intelligent flower thinning in pear orchards,and adopts 3D SLAM method to construct a map of the scaffold pear orchard environment,and uses RTK and NDT point cloud matching for positioning,so that the inspection platform has the function of positioning itself against each ranks in the orchard.The orchard navigation path planning is realized by annotating point cloud high precision maps.Based on computer vision technology,the fast recognition function of pear tree inflorescence by the navigation platform is realized by improving YOLOv5 algorithm.The main research contents of this article are as follows:(1)Software and hardware construction of autonomous inspection and target recognition test platform.On hardware,the map perception and acquisition system is composed by fusing orchard vehicle motion chassis,multi-line LIDAR and inertial measurement unit;pear tree inflorescence image data is collected by depth camera.On the software,the framework of vehicle navigation and inflorescence recognition program is built based on Linux system;the ROS operating system is used to build the orchard 3D environment and complete the functions of map repositioning and motion control;the Pytorch library is used to realize the training of pear tree inflorescence deep learning model.(2)Design of an autonomous navigation system for orchards based on 3D SLAM.The point cloud information of orchard is collected by 3D Li DAR with NDT SLAM algorithm,and the filtering method of PCL point cloud library is used to extract the trellis point cloud as the matching target.The test results show that the error of the trellis pear orchard point cloud is less than 3 cm,which meets the subsequent point cloud positioning requirements.In terms of positioning method,by transforming RTK information into vehicle initial position information,the frame point cloud and trellis point cloud are transformed with normal distribution to determine the real-time vehicle position.In terms of path planning,the navigation path is specified by manually planning the vector map,and the trajectory tracking is realized by pure path tracking controller.(3)Design of YOLOv5 based pear tree inflorescence recognition system.The images of pear tree buds and flowers are collected for annotation and data expansion,and the improved YOLOv5 algorithm is used for deep learning training and obtaining detection models.To address the problem of dense pear tree inflorescences and low recognition efficiency of the original model,the network’s recognition accuracy of the target is improved by adding CA attention mechanism.Using Ghost module to replace the traditional convolutional layer greatly reduces the number of model parameters without reducing the model accuracy.The improved YOLOv5 model achieves an accuracy of 91.3% and a recognition speed of 23 ms.Compared with the original YOLOv5 model,the accuracy and recall of the improved model are increased by 1.9% and 3.1%,respectively,and the number of parameters and model size are reduced by 46.6% and 47.8%,respectively,which improves the operating efficiency of the device.The average detection time was reduced by 4 ms.(4)Field experiments of variable speed and fixed-point navigation and pear inflorescence recognition in pear orchards.The variable-speed and fixed-point navigation experiments were carried out in a real trellised pear orchard field environment,and the results showed that the navigation platform worked best when running at a speed of 1 m/s,when the lateral positioning error of RTK fused with NDT point cloud matching was within 5 cm and the heading positioning error was within 1°.In the inflorescence identification test,by comparing with manual measurement,the inflorescence identification method of this study has an accuracy rate as high as80.98%,which meets the intelligent flower thinning requirements of the subsequent pear orchard.