Detecting Defensive Honeybee Behavior at the Hive Entrance: Benchmarking YOLO Models on an Embedded CPU Platform

Defensive behavior among honeybees at the hive entrance, such as colony defense against intruders, appears as a group-level pattern rather than in individual bees. This study investigates its automatic recognition from single frames as a single-class object-detection task suitable for edge deployment. A dataset of 20,000 annotated images was built from 87 videos (13 h) recorded at 20 hive entrances; defense was observed at 6 entrances in 32 instances totaling about 6 min. YOLOv8, YOLO11, and YOLO26 detectors at five scales were evaluated by five-fold cross-validation, with folds grouped by defense instance, and exported to ONNX and NCNN for CPU-only benchmarking on a Raspberry Pi 5. The models reached ${mAP}_{50}$ of 0.728–0.800 and ${mAP}_{50-95}$ of 0.447–0.541, with YOLO11 and YOLO26 generally more accurate than YOLOv8, although not in every split. Larger models improved ${mAP}_{50-95}$ by 4.6–6.4 percentage points, which is consistent across all splits but small relative to the 14–20× increase in time per frame, and export changed accuracy by, at most, 2.5 percentage points on average. NCNN was the fastest and most energy-efficient format, with nano-models reaching about 13–14.5 FPS at 0.4–0.5 J per frame. YOLO11n, YOLO26n and YOLO11s appear to offer the best frame-level accuracy–speed trade-off; generalization beyond one apiary and event-level performance remain to be verified.