Sledevič Tomyslav, Serackis Artūras, Matuzevičius Dalius, Plonis Darius, Vdoviak Gabriela
Department of Electronic Systems, Faculty of Electronics, Vilnius Gediminas Technical University - VILNIUS TECH, Vilnius, Lithuania.
PLoS One. 2025 Feb 25;20(2):e0318401. doi: 10.1371/journal.pone.0318401. eCollection 2025.
This study presents a novel method for automatically recognizing honeybee behavior patterns at the hive entrance, significantly contributing to beekeeping and hive management. Utilizing advanced YOLOv8 models for detection and segmentation, our approach analyzes various aspects of bee behavior, including location, direction, path trajectory, and movement speed within a designated area on the hive's landing board. The system effectively detects multiple bee activities such as foraging, fanning, washboarding, and defense, achieving a mean detection accuracy of 98% and operating at speeds of up to 36 fps, surpassing state-of-the-art methods in both speed and accuracy. Key contributions include the development of a comprehensive dataset with 7200 frames from eight beehives, the introduction of the first known research focused on recognizing bee behavior patterns through visual analysis at the hive entrance, and a comparative evaluation of various object detection and tracking algorithms tailored for bee detection and behavior recognition. Our findings indicate that this method enhances monitoring capabilities for beekeepers while reducing the need for manual inspections, thereby minimizing disturbances to the bees. By analyzing spatial trajectories and occurrence density maps, the proposed framework provides robust identification of overlapping behaviors, facilitating timely interventions when necessary. This work lays the groundwork for future automated monitoring systems aimed at improving hive health and productivity.
本研究提出了一种在蜂巢入口自动识别蜜蜂行为模式的新方法，对养蜂和蜂巢管理有显著贡献。利用先进的YOLOv8模型进行检测和分割，我们的方法分析蜜蜂行为的各个方面，包括在蜂巢着陆板指定区域内的位置、方向、路径轨迹和移动速度。该系统有效地检测出多种蜜蜂活动，如觅食、扇风、搓板行为和防御行为，平均检测准确率达到98%，运行速度高达36帧/秒，在速度和准确率方面均超越了现有方法。主要贡献包括开发了一个包含来自八个蜂箱的7200帧图像的综合数据集，引入了首个专注于通过蜂巢入口视觉分析识别蜜蜂行为模式的已知研究，以及对为蜜蜂检测和行为识别量身定制的各种目标检测和跟踪算法进行了比较评估。我们的研究结果表明，该方法增强了养蜂人的监测能力，同时减少了人工检查的需求，从而最大限度地减少了对蜜蜂的干扰。通过分析空间轨迹和出现密度图，所提出的框架能够对重叠行为进行可靠识别，便于在必要时及时进行干预。这项工作为未来旨在改善蜂巢健康和生产力的自动化监测系统奠定了基础。
