Automatic annotation rules for stem orientation in support-based mechanical harvesting of ‘Yuluxiang’ pear

As a key prerequisite for perception systems in low-damage mechanical harvesting of ‘Yuluxiang’ pear, stem-orientation acquisition is essential for calculating lifting direction and ensuring harvesting stability. Under natural orchard conditions, the small size and frequent occlusion of pear stems make orientation annotation labor-intensive, rendering reliable stem-orientation representation and efficient dataset construction challenging. To address these issues, this study proposed an automatic annotation framework for stem-orientation acquisition. Using manual annotations as the reference, two directional annotation forms, keypoint-based pose (Pose) and oriented bounding box (OBB), were designed and compared. Based on their different orientation representation mechanisms, method-specific geometric consistency criteria and one-to-one matching rules were formulated, with base-point distance d b a s e and angular deviation Δ θ used for Pose, and area overlap ratio k and angular deviation used for OBB. After YOLO-based benchmark model selection, confidence threshold scanning, and decision parameter analysis, the final joint constraints were determined as d b a s e ≤ 6 px and Δ θ ≤ 10 ∘ for Pose, and k ≥ 0.5 and Δ θ ≤ 10 ∘ for OBB. Test set evaluation showed comparable automatic annotation performance between Pose and OBB, and the successfully matched instances were further analyzed to compare orientation accuracy. The mean and median angular deviations of Pose were 3.313° and 2.748°, respectively, lower than those of OBB, which were 3.719° and 3.140°. The proportions of matched instances with angular deviations not exceeding 3° and 5° were 54.22% and 74.77% for Pose, compared with 48.55% and 69.11% for OBB. These results indicate that Pose is preferred for stem-orientation dataset construction and lifting direction calculation.