Cooperative lane mapping for autonomous mobile robots in horticulture by aerial photogrammetry

摘要

• Lane paths, 3D-, obstacle grid-, and structure maps are derived from UAS data to ease robot deployment. • Structure maps enable path planning even in overgrown lanes or sections where crops are missing. • The method demonstrated high accuracy and generalizability across diverse orchards and crops. • It is computationally efficient and ROS2-integrated. • The generated maps have potential use in robot localization. Classical methods for navigation of autonomous mobile robots often fail in orchards with dense inter-row vegetation such as tall grass and protruding branches. Since robustness and ease of use are essential to the adoption of innovative technology, this can be a hindrance for the development of new types of robotic services. At the same time, the use of unmanned aerial systems for monitoring applications in horticulture is increasing. In this study we analyse whether this data can be used to derive navigation maps and graphs of orchards, even in conditions where classical navigation methods fail. To answer this question, we developed and analysed an automated approach that derives maps with distinct levels of detail from unmanned aerial system imagery for robot navigation in orchards. We test our approach on an apple tree orchard with missing plants, and a mixed-berry orchard with strong inter-row vegetation and overarching branches. We generate voxel maps, 2D grid maps and topological maps and introduce the concept of structure maps. These enhance the existing row structure of the orchards. We show that our method for grid map computation generalises to different crop types without retraining, while providing grid maps that are on-par with, or better than current machine learning based approaches with mean IoUs of 0.87 to 0.97. While the structure maps show limitations in areas without any trees or plants, the derived paths show mean absolute errors equal to or less than 0.12 m compared to GNSS-based references, which is accurate enough for many agricultural applications.