Blueberry Row Detection Based on UAV Images for Inferring the Allowed UGV Path in the Field

Abstract

Blueberries (Vaccinium corymbosum) have become one of the most popular fruits in Europe, being ranked as the top-berry fruit species. Due to the high cost and poor availability of the workforce, agricultural production of blueberries is becoming heavily reliant on robots. Within the EU-supported Flexigrobots project, unmanned ground vehicle (UGV) solutions are being developed for weeding, soil analysis, and spraying. However, to detect the problematic regions (regions of interest) and plan the UGV trajectories, fields are first scanned using unmanned areal vehicles (UAVs). In this paper, we propose a procedure for processing UAV images for inferring the allowed UGV trajectories in blueberry fields. First, the image dataset was acquired using a UAV equipped with a multispectral camera. Afterward, we utilized U-Net-based deep neural network (DNN) models for semantic segmentation of blueberry rows and enhanced the results with Radon transform and additional image post-processing steps. The quality of the obtained segmentation mask is analyzed through the ranking of trained models in terms of the intersection over union (IoU) and F1-score metrics. Lastly, based on the detected rows and the inter-row regions, we generated the trajectory along which the UGV is allowed to move to perform the weeding, soil analysis, or spraying tasks.