Cutting Point Detection for Strawberry Fruit Harvesting and Truss Pruning by Agricultural Robot*

This study aims to develop an agricultural robot with multiple functions. Moreover, the study focuses on the functions of harvesting and pruning and proposes a method for detecting cutting point to achieve these functions. This method comprises three processes. First, the input image was classified into five classes using semantic segmentation via deep learning. Next, properties in fruit, calyx, and truss were extracted. Finally, based on those properties, cutting points for fruit harvesting and truss pruning were detected. This method detected the cutting points based on plant features. Additionally, the unworkable area for the robot was estimated, and the possibility of cutting was determined. The method was evaluated using 1,000 images acquired in a strawberry greenhouse. The accuracy for harvesting and pruning were 0.93 and 0.86 in the cutting point detection and 0.96 and 0.88 in the mature and immature fruits detection, respectively. This paper discussed the causes of false and missing detections and their solutions.