Development of Smart Viticulture System Using Autonomous Robots

该研究旨在通过使用自主机器人和人工智能开发智能葡萄栽培系统。智能葡萄栽培旨在通过自动化葡萄园监控、检查、产量估算和收获葡萄等任务，减轻农民的负担。在2022财年，通过使用基于深度学习的地标检测，提高了上一财年开发的监测系统的准确性。以前开发的系统依赖于基于颜色的地标（柱）检测，这在光照条件和降雨的变化中导致不准确。对基于卷积神经网络的模型进行了训练和评估，使其对环境的动态变化具有鲁棒性，并能更好地估计地标。这改进了整体SLAM（同步定位和地图）模块，该模块的开发不使用GPS，并且依赖于仅本地特征以降低成本。此外，还开发了一个深度学习CNN模型，用于检测葡萄/树枝及其状态（成熟、快成熟、未成熟）。开发了一种杂草检测模型。收获模块是在模拟器中开发的，计划在2023财年进行真实世界的制造和测试。结构化葡萄园环境中的机器人导航算法、避障、报警系统、葡萄园多机器人合作算法等也得到了开发，并在会议上发表了成果。

This research aims at developing a smart viticulture system through the use of autonomous robots and AI. Smart viticulture aims at reducing the burden of farmers by automating tasks like vineyard monitoring, inspection, yield estimation, and harvesting of grapes.In FY2022, the accuracy of the monitoring system developed in previous FY was improved through the use of Deep Learning based landmark detection. The previously developed system relied on landmark (pillar) detection based on color, which caused inaccuracies during the change of illumination conditions and rain. A convolutional neural network based model was trained and evaluated to be robust against dynamic changes in the environment and gave better estimation of the landmarks. This improved the overall SLAM (Simultaneous Localization and Mapping) module which was developed without the use of GPS, and relies on local-only features for cost-cutting. In Addition, a deep learning CNN model was developed for detecting grapes/branches and their conditions (ripe, almost-ripe, raw). A weed detection model has been developed. The harvesting module is developed in simulator and a real-world manufacturing and test is planned for FY2023. Algorithms for robot navigation in structured vineyard environment, obstacle avoidance, alarm system, multi-robot cooperation algorithms for vineyard, have also been developed and results have been published in conferences.