Bee Detection and Tracking at Hive Entrance using YOLO11 and ByteTrack

This work combines YOLO11 transfer learning with ByteTrack for automatic bee entrance monitoring. It compares data augmentation, backbone freezing, and tracker settings for small, fast-moving bees. Progressive backbone unfreezing achieves approximately 97.0 percent precision and 98.7 percent mAP50, while light augmentation performs better than heavy augmentation. On an independent 25 FPS side-view video, the optimized system correctly counts 43 of 47 incoming bees (91.5 percent) but only 7 of 30 outgoing bees (23.3 percent). Error analysis attributes most counting errors to missed detections caused by rapid motion and blur, although tracker optimization reduces tracking failures. The results demonstrate why strong frame-level detection metrics must be supplemented by directional counting evaluation.

Publication details

Authors
Thi Thu Thao Nguyen, Johannes Reschke
Organizations
🇫🇮 Savonia University of Applied Sciences🇩🇪 Ostbayerische Technische Hochschule Regensburg
Year
2026
Type
Preprint

Publication and access

The abstract above is an editorial summary of the paper. Affiliations are taken from the PDF's first page.

Relevancy to Gratheon

The detector-plus-tracker architecture closely matches an entrance-observer pipeline. Moderate augmentation, progressive unfreezing, and low-confidence association are useful implementation hypotheses for Gratheon's own videos.

The main product lesson is the large directional performance gap. An apparently strong detector can still severely undercount departing bees, making an inferred return ratio or colony-loss alert misleading. Gratheon should evaluate incoming and outgoing count errors separately and test exposure time, frame rate, motion blur, and camera geometry before interpreting those counts as colony-health signals. A single independent video is not evidence of cross-apiary reliability.