Kyoho grape picking localization method based on improved YOLOv8n-pose model

Objective To accurately and efficiently localize the picking position of Kyoho grapes so as to effectively reduce fruit damage.

Method A grape picking localization method based on improved YOLOv8n-pose was proposed. Firstly, the improved YOLOv8n-pose was utilized to detect keypoints of the grape stem and the vulnerable grapes at the top. Based on the coordinates of these keypoints, a characteristic vector representing the upper boundary pose of the grapes was constructed. This vector was then used to calculate the optimal picking angle. Finally, the optimal picking position was determined through the synergy of the picking point and picking angle.

Result Experimental results showed that the precision (P), recall (R), mAP@0.50 and mAP@0.50~0.95 of the improved YOLOv8n-pose increased by 1.7, 0.7, 0.9 and 1.7 percentage points respectively compared to the original model, and increased by 0.4, 0.1, 0.6 and 2.7 percentage points respectively compared to YOLOv12s-pose. Meanwhile, the number of model parameters was reduced by 5.8% compared to YOLOv8n-pose. The successful localization rate using the proposed method reached 90.8%, which was an improvement of 9.2 percentage points over methods that did not use the picking angle.

Conclusion This study provides a low-damage picking localization method for Kyoho grape harvesting robots.