Accurate and robust pollinations for watermelons using intelligence guided visual servoing

With a significant decline in the bee population, there is an increasing demand for automated robotic pollination. This study proposes a novel approach for automating watermelon pollination using visual intelligence-guided servo control. On the control loop, the sizes and orientations of flowers are estimated by leveraging the inference capability of Deep Learning. The estimated sizes of flowers are then converted to their corresponding depth information, which can be utilized to determine their coordinates afterward. Thus, visual intelligence serves as an essential element in the control loops, namely, intelligence-guided visual servoing. One of the promising features of the proposed depth measurement is that its measurement accuracy becomes higher at a distance closer than 100 mm. The robustness of the proposed method at a closer distance has been verified through the depth sensitivity analysis of the size estimation error. Leveraging a newly compiled watermelon flower dataset, the study conducted over 50 experiments in open-field watermelon cultivation. The achieved results showcase a high detection rate, yielding a mean average precision (mAP) of 90.9 %. The average depth error associated with pollination target localization was a mere 1.028 cm, while the pollination speed remained an average of 8 s per watermelon flower, underscoring its practical feasibility.