Reliable Visual Guidance for Robotic Safflower Harvesting Using Corolla–Bud Topology and Selective Dual-Camera Correction

Field safflower harvesting is challenged by target-scale variation, occlusion, and random growth orientations, which can degrade visual perception and end-effector alignment. This study proposes a robotic visual-guidance framework integrating FAPB-Net segmentation, corolla–bud topology, and selective main-auxiliary dual-camera correction. Inclination is estimated from the fitted corolla ellipse, while in-plane orientation is obtained from the directed bud-to-corolla vector. Auxiliary inclination estimates are introduced only when three-dimensional target matching and angular consistency are satisfied. FAPB-Net achieved corolla and bud intersection over union (IOU) values of 90.28% and 70.85% at 28 FPS. On 138 synchronized dual-camera test pairs, inclination mean absolute error (MAE) decreased from 7.5° to 6.3°, while Acc@10°, defined as the proportion of estimates with absolute error below 10°, increased from 86.2% to 90.6%. In field tests with 515 targets, recognition/localization and end-effector arrival rates reached 95.92% and 94.95%, respectively. For 447 targets with complete reference measurements, inclination and orientation MAEs were 6.0° and 4.72°. Pose-guided harvesting achieved a mean harvesting completeness rate (HCR) of 94.63%, compared with 91.40% for position-only guidance. These results indicate the practical value of pose information for reliable safflower harvesting.