Research on robot vision detection and yield estimation of citrus fruits at maturity based on improved YOLOv5

Abstract

To address low citrus detection accuracy caused by small fruit size, occlusion and overlap, as well as the neglect of inter-tree growth differences in single-plant yield estimation, this study proposes an optimized YOLOv5 algorithm for mature citrus fruit detection. A yield estimation model for round-headed Newhall navel oranges was established based on single-sided canopy recognition, with a fitted coefficient of determination R² of 0.6034 between the fruit yield identified from unilateral canopy images and the actual per-plant fruit yield. The yield prediction model was deployed on a robot inspection vehicle. Taking Orah and Newhall navel oranges as examples, tests confirmed that yield estimation models for different varieties cannot be used universally. Field validation showed that the overall relative error of the yield estimation process was only 5.2%, with significantly higher efficiency than manual statistics. The proposed mature citrus detection and yield estimation method can provide a basis for formulating scientific, reasonable and personalized precision fertilization plans for fruit trees in the coming year.