Growth-Point Detection and 3D–2D Two-Stage Localization of Cucurbit Rootstock Seedlings Using YOLO11-Pose-LGA

In automated grafting of cucurbits, variation in growth-point position and morphology of cucurbit rootstock seedlings, together with position deviation caused by picking and grasping, can limit the millimeter-level accuracy required for growth-point cutting. This study proposed a 3D and 2D two-stage visual localization method based on the lightweight YOLO11-Pose-LGA model. In the first stage, RGB-D depth information was used to construct a 3D point set and estimate the 3D position and grasp pose of rootstock seedlings in random poses, thereby guiding adaptive robotic picking. In the second stage, side-view 2D vision was used to fit the V-shaped inner edges of the cotyledons after grasping, relocalize the growth point, and compensate for post-grasp position deviation before cutting. The YOLO11-Pose-LGA model achieved a Pose mAP@0.5:0.95 of 0.916. In physical system tests involving 400 rootstock seedlings, 378 seedlings were successfully picked, corresponding to a 3D vision-guided picking success rate of 94.5% (378/400). The second-stage 2D growth-point relocalization for the 378 successfully picked seedlings produced a vertical MAE of 0.39 mm and a combined MAE of 0.63 mm. After growth-point relocalization and Z-axis height compensation, 342 of the 378 successfully picked seedlings met the cutting qualification criteria, corresponding to a conditional qualified cutting rate of 90.5% (342/378) and an end-to-end successful picking-and-cutting rate of 85.5% (342/400). These results indicate that, under the tested experimental conditions, the proposed 3D global guidance and 2D local compensation framework can support spatial picking guidance for rootstock seedlings in random poses and post-grasp pre-cut growth-point localization for automated grafting operations.