Computer vision in precision poultry farming: evaluation of YOLO and transformer-based object detectors for dead broiler detection in farm inspection images
Veterinary and Animal Science · 2026-12 · 34 · 100923
DOI: 10.1016/j.vas.2026.100923 ↗Kurz gefasst
Fachartikel in der Zeitschrift Veterinary and Animal Science, 2026-12. Verfasst von Sandra Nicole Umurungi und 9 weiteren Autor:innen. Beteiligte Einrichtung: Department of Technology and Cybernetics, Faculty of Agriculture and Technology, University of South Bohemia in Ceske Budejovice, Studentska 1668, 37…. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Technology Agency of the Czech Republic.
Abstract (Originalsprache)
Broiler mortality inspection is an important component of poultry farm management. This study presents a pilot technical feasibility evaluation of recent YOLO and transformer-based object detection models for detecting dead broilers in event-enriched farm inspection images. The dataset was collected from a commercial farm over a 33-day production cycle and contained 1068 images with 7222 live and 1016 dead broiler instances. Images were captured after mortality events had been identified during routine inspection and before carcass removal. Six models, YOLOv12-L, YOLOv12-X, YOLO26-L, YOLO26-X, RT-DETR-L, and RT-DETR-X, were evaluated under a shared RGB-based protocol across three independently trained seeds. Dead broiler precision, recall, mAP50, and mAP50-95 were defined as the primary performance endpoints, while all-class mAP50-95, training stability, and computational efficiency were considered secondary descriptive analyses relevant to future decision-support development. YOLO26-L achieved the highest dead-class mAP50-95 (0.912 ± 0.003) and all-class mAP50-95 (0.875 ± 0.002), while RT-DETR-L achieved the highest dead broiler recall (0.972 ± 0.001). Differences among leading models were small and were interpreted descriptively rather than as evidence of model superiority. Under the A100 evaluation protocol, YOLO26-L also showed low inference time (2.23 ± 0.06 ms/image). Overall, the findings support the technical feasibility of image-based dead broiler detection in mortality-containing images. The combined analysis of accuracy, multi-seed training stability, and computational efficiency further provides practical insights for future image-based poultry mortality inspection systems. However, future studies should include unbiased negative scene sampling, event-level identifiers, external farm validation, deployment hardware, and outcome-level evaluation of inspection efficiency and welfare-related benefits.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Department of Technology and Cybernetics, Faculty of Agriculture and Technology, University of South Bohemia in Ceske Budejovice, Studentska 1668, 37005 Ceske Budejovice, Czech Republic.
- Institute of Industrial Electronics, Electrical Engineering and Energy, Riga Technical University, LV-1048 Riga, Latvia.
Förderung
- Technology Agency of the Czech Republic
Zitieren
APA
Umurungi, S. N., Ufitikirezi, J. D. M., Bumbálek, R., Zoubek, T., Kuneš, R., Stehlík, R., Havelka, Z., Kalimu, K., Avotins, A., & Bartoš, P. (2026). Computer vision in precision poultry farming: evaluation of YOLO and transformer-based object detectors for dead broiler detection in farm inspection images. Veterinary and Animal Science, 34, 100923. https://doi.org/10.1016/j.vas.2026.100923
Vancouver
Umurungi SN, Ufitikirezi JDM, Bumbálek R, Zoubek T, Kuneš R, Stehlík R, et al. Computer vision in precision poultry farming: evaluation of YOLO and transformer-based object detectors for dead broiler detection in farm inspection images. Veterinary and Animal Science. 2026;34:100923. doi:10.1016/j.vas.2026.100923
BibTeX
@article{Umurungi2026,
author = {Umurungi, Sandra Nicole and Ufitikirezi, Jean de Dieu Marcel and Bumbálek, Roman and Zoubek, Tomáš and Kuneš, Radim and Stehlík, Radim and Havelka, Zbyněk and Kalimu, Karimunda and Avotins, Ansis and Bartoš, Petr},
title = {Computer vision in precision poultry farming: evaluation of YOLO and transformer-based object detectors for dead broiler detection in farm inspection images},
journal = {Veterinary and Animal Science},
year = {2026},
volume = {34},
pages = {100923},
doi = {10.1016/j.vas.2026.100923}
}
EUVIMED