Chang, Han BinHan BinChangChang, Shan ChengShan ChengChangChueh, Cheng YuCheng YuChuehLin, Hung JenHung JenLinJOE-AIR JIANGLiu, An ChiAn ChiLiuEN-CHENG YANGHsieh, Hsiang WenHsiang WenHsiehCHENG-YING CHOU2024-02-192024-02-192023-01-019781713885887https://www.scopus.com/inward/record.uri?eid=2-s2.0-85183583397&doi=10.13031%2faim.202300602&partnerID=40&md5=38fe93159557ea71edda5204c7066d69https://scholars.lib.ntu.edu.tw/handle/123456789/639775The pollination of facility cultivation is heavily dependent on honeybees. In limited space, using too many beehives can result in wasted costs and overcrowding, which can negatively impact the health of the bees due to increased competition. Conversely, using too few beehives can lead to a low pollination rate. Therefore, it is essential to identify an effective method for evaluating honeybee pollination rates. To address this issue, this research aims to count the number of honeybees entering and leaving the nest, as well as identify the pollen grains collected by the bees, to evaluate their pollination rate. Raspberry Pi HQ cameras have been installed on the lower, left, and right side of an acrylic observation channel to capture videos of bees entering and leaving the nest. Additionally, we used the YOLOv5 object detection model with an s6 backbone for honeybee recognition, achieving an accuracy of 99.6%. To prevent counting errors, the model's identification results were input into the StrongSORT tracking algorithm(s). By tracking the start and end points and length of the bees' trajectories, we set counting rules for entering and leaving the nest. As a result, the number of bees entering and leaving the nest can be determined. Furthermore, we extracted images of bees entering the nest from both sides and input them into the YOLACT instance segmentation model and Swin vision transformer separately for identification and comparison. The prediction accuracy of the two models in the pollen grain mask exceeded 86%. By combining the counting of bees entering and leaving the nest with the identification of pollen grains, we established an evaluation index for bee pollination rate. This allows for more accurate and efficient monitoring of the pollination process. In addition to knowing the number of bees entering and leaving the nest, as well as the collected pollen grains, it is possible to better determine the pollination rate and make necessary adjustments to increase crop yields and reduce costs.honeybee | image segmentation | object detection | object trackingconference paper10.13031/aim.2023006022-s2.0-85183583397https://api.elsevier.com/content/abstract/scopus_id/85183583397