Résumé
Reliable honey bee detection in hive images remainschallenging, as bees are small, visually similar and often affected by overlap, occlusion and uneven illumination. This paperpresents a real-time hybrid detection framework for infraredhive scenes, merging an image-processing detector with a YOLObased detector. The image-processing detector operates in realtime on a 30 fps video stream, combining classical imageprocessing algorithms to identify bees from brightness contrastand body-shape information. In parallel, several medium-scaleYOLO models are trained and compared on a dedicated tileddataset prepared from annotated hive images. Among the evaluated models, YOLOv8m achieved the highest precision, recall andmAP50-95, with scores of 0.92, 0.92 and 0.69, respectively, andwas therefore selected for the hybrid stage. The final frameworkmerges the detections from both methods using overlap andcontainment criteria, preserving complementary detections whilelimiting duplicates. Experimental results show that the hybriddetector improves bee detection coverage, especially in denseand partially occluded regions where a single detector maymiss visible bees. The proposed framework provides a practicalstep towards real-time hive monitoring and future behaviouralanalysis.