Pedicel localization method for pomelo harvesting robots in unstructured orchard environments

Precise localization of pedicel picking points is critical for automated pomelo harvesting robots. However, the tall canopy structure of pomelo trees, highly randomized spatial distribution of fruits, and dense foliage pose significant challenges for accurate pedicel positioning in unstructured orchard environments. This study proposes a depth map-based method for calculating pomelo pedicel picking points to achieve high-precision localization under complex field conditions.First, an unstructured orchard scene pomelo dataset comprising 5,706 samples was constructed. Based on RGB-D imaging and YOLOv8-seg-RepGhost, segmented depth maps of pomelo fruits were extracted. Through RGB-D registration and semantic mask alignment, a depth-geometric collaborative analysis framework was established, achieving sub-centimeter level measurements of both longitudinal and transverse fruit diameters. Second, a novel concept of "pose feature-sensitive region" located at the transition zone among the fruit shoulder, equator, and base surface was proposed. The principal axis direction of isodepth contours within this feature-sensitive region was dynamically extracted, and Hough ellipse detection was extended to the depth feature space. Finally, a geometric mapping model between fruit pose and pedicel position was constructed, enabling spatial coordinate calculation for occluded pomelo pedicels. Experiments demonstrate that within the common 50–150 cm working range of robotic arms, the average longitudinal diameter error ranges from 0.385 to 0.675 cm, while the average transverse diameter error ranges from 0.407 to 0.724 cm. In unobstructed scenarios, fruit pedicel positioning error do not exceed 2.5 cm with angular deviation below 6°. Under severe occlusion, positioning error remained within 3 cm, with an average angular error of approximately 3.7°. This study effectively enhances pedicel positioning accuracy and robustness in complex orchard environments, providing theoretical support and technical references for precise harvesting operations in unstructured settings.