Fig 1.
Data acquisition on local apiary.
(1) Beehive, (2) portable camera for dataset collection at the hive entrance, (3) beehive landing board.
Fig 2.
Conceptual algorithm of the bee behavior pattern recognition on the hive landing board.
Fig 3.
Images of 8 beehive entrances with annotated bees in the publicly provided bee detection dataset [10].
Images (a–h) represent different entrance ramps. Five hives (d–h) feature extended landing boards, used to facilitate bee landings. All the bees are assigned to one class, regardless of whether they are partially occluded, fully visible, blurred, or carrying pollen. Drones (g) are classified as bees. Number of annotated bees at the entrances of 8 beehives (i).
Fig 4.
Triangle-annotated bees using the Labelme tool for direction estimation (a–e).
Segmented bee contours (in red), minimum area rectangle (in blue), and bee direction vectors (in green) (f–j).
Fig 5.
Conceptual algorithm of bee direction vector estimation based on segmentation.
Fig 6.
The principle of defense pattern detection (a).
The conceptual algorithm for defense detection (b). The pink circle (c) marks the group of the thief and defenders after processing single frame with the defense detection algorithm. The red circle marks the defense pattern after post-processing the defense detections (d).
Fig 7.
The representative 2D trajectories of honeybees while defending (a–b), the 2D distances in the x and y directions vs. time (c–d).
The thief and two guard bees participated in the defense for the first 20 s, and for the next 45 s, during the period from 25 s to 70 s (c–d).
Fig 8.
Speed of motion patterns on the hive landing board while foraging (a).
Conceptual algorithm for foraging detection (b).
Fig 9.
Speed of motion patterns on the hive landing board while fanning (a).
Conceptual algorithm for fanning detection (b).
Fig 10.
Speed of motion patterns on the hive landing board while washboarding (a).
Conceptual algorithm for detecting washboard movement (b).
Table 1.
Parameters used in the activity recognition algorithms.
Table 2.
Comparative evaluation of the proposed implementation to state-of-the-art methods for bee detection in images.
Table 3.
Training results of the YOLOv8 models used in bee detection and direction estimation.
Table 4.
Comparison of tracker accuracy and update time across different tracking algorithms applied to 1920 × 1080 video with a detector confidence threshold of 0.3. Tracker update time is measured in milliseconds per frame.
Table 5.
Average time per algorithm stage. In the first line, YOLOv8m is used for bee detection in 1920 × 1080 px frames, and YOLOv8n-seg is applied for bee segmentation in 64 × 64 px cropped images. In the second line, YOLOv8m-seg is used for both bee detection and segmentation.
Fig 11.
Occurrence density maps of the bee tracks on the landing board after 5 s (a), 10 s (b), and 20 s (c) of monitoring the entrance zone.
Fig 12.
Detected bees on the landing boards.
Foraging bees are annotated in yellow (a–c), with green tracks showing the paths of the bees during the last second. The pink circle marks the detected defense event (c, d). Fanning bees are marked with cyan bounding boxes (a, b, e). Bees exhibiting washboard movement are annotated in orange (f), and occurrence density maps of the last two seconds are captured within white polygon-constrained zones (e, f). Optional direction vectors are marked in yellow (d, e).
Fig 13.
The distribution of the activities in the behavior classification dataset. Horizontal axis presents frame index, vertical axis shows number of tracked bees assigned to one of the four behavior classes. Vertical lines separate five different records. The corresponding annotated frames with detected activities in these records are presented in Fig 12a, 12b, 12c, 12e, 12f, respectively.
Fig 14.
Confusion matrix for bee behavior pattern recognition.