トマト果実収穫ロボットのための障害物領域抽出機能の開発と評価

抄録

In this paper, we developed and evaluated an obstacle area extraction function for a robot that aims to automate agricultural work in the cultivation environment of mini tomatoes. We proposed a process flow for determining harvestable fruit by considering fruit positions obtained via object detection on camera images and obstacle areas obtained through semantic segmentation. Our proposed method applies DeepLab v3+ model to segment the images into each class and performs post-processing such as removing misclassified areas and removing areas using a depth camera. We evaluated the semantic segmentation task with four classes, including main axis, main axis bundle, pillar, and background, or five classes including hose, using mean IoU and accuracy as evaluation metrics in the learning experiment. The proposed method is expected to contribute to the evaluation of obstacle detection functions, one of the challenges for practical application of harvesting robots.