Oblate Rotational Ellipsoid-Based Spatial Pose Estimation of Blueberry Fruits Using an Improved YOLOv11 Multi-Task Perception Network

Accurate 3D (three-dimensional) pose perception of blueberries is essential for non-destructive twist picking by robotic harvesters. While recent studies on fruit pose estimation have achieved promising results, most existing methods rely on a sphere assumption that introduces inherent systematic errors for naturally oblate fruits such as blueberries. Our previous work developed a lightweight YOLOv11 (the eleventh version of You Only Look Once) detector and estimated the 3D pose of the fruit growth axis under the standard sphere assumption. To address this geometric limitation, this paper innovatively presents an oblate-ellipsoid projection model matching blueberry morphology, together with an ellipse major-axis method that eliminates the systematic errors caused by the standard sphere approximation in angle computation. In addition, the original detector is upgraded to a multi-task perception network that jointly performs fruit detection, calyx keypoint localization, and pixel-level contour extraction, providing comprehensive information for subsequent blueberry pose analysis. The experimental results show that the improved network achieves 95.7% detection mAP (mean average precision), 86.5% keypoint mAP, and 78.4% segmentation mAP, with only 7.8% higher computational complexity than the original detector. The average angular error between the fruit growth axis and the camera optical axis drops from 19.2° to 10.9°, and the error at the ideal picking angle decreases from 10.0° to 4.7°, without sacrificing real-time performance. This method overcomes the geometric limitation of the original approach and provides more reliable visual support for non-destructive blueberry picking.