Multi-vision-based Picking Point Localisation of Target Fruit for Harvesting Robots
To address inaccurate fruit harvesting-point localization—leading to fruit damage and drop—in agricultural harvesting robots, this paper proposes a dual-path localization method integrating multi-view RGB-D vision with model-driven reasoning. Innovatively, motion-capture calibration is fused with surface points from two calibrated cameras (Cfix/Ceih) to construct a multi-view geometric model; furthermore, AdaBoost regression is first introduced and empirically validated for harvesting-point prediction, demonstrating superior robustness over conventional single-camera analytical methods. Experiments show that the proposed method achieves an 88.8% harvesting success rate (median localization error: 4.40 mm), outperforming the single-camera baseline by 11.1 percentage points and significantly improving overall robotic harvesting performance. Key contributions include: (i) the first multi-view geometric modeling framework tailored for fruit harvesting-point localization, and (ii) the first application and validation of AdaBoost regression for this task, establishing its superiority in accuracy and robustness.