Robust Visual Servoing for Precision Agriculture Tasks

Recently, robots started making their first steps towards real-world applications in agriculture and more specifically, in vineyards. Among other challenges, recognizing clusters of grapes and performing visual servoing towards them is an important task. Although approaches such as deep learning have emerged that seem to simplify the problem, and databases for training data are publicly available, results are affected severely by weather conditions. In this paper, the detection robustness of grape clusters is investigated subject to rain conditions, with the use of two state-of-the-art models for object detection, Mask-RCNN, and YOLOv3. It is shown that rain in an image markedly reduces the accuracy of the classifiers, indicating that a de-raining method is vital in classification and training detection methods with rainy images is not enough. Cycle-GANs are exploited to generate de-rained images from rainy samples. The method is validated in a lab experiment using a wheeled robotic platform and a low-cost onboard computer. Mask-RCNN proves to be computationally intensive to run onboard compared to YOLOv3. In this scope, we demonstrate a complete, robust under rainy weather, low-cost, and expandable application for precision agriculture in which a robot identifies a cluster of grapes at a high frequency by running YOLOv3-tiny on board and approaches it at a predefined distance.