GDLYOLO-Tracker: Real-time multi-target tracking of poultry open-mouth behavior in high-density houses under nighttime infrared conditions

Read full text Summarise with VIRAT
Authors: Zeyang Bi, Liwei Yang, Man Zhang, Han Li
Cited by: 0
References: 33

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Poultry respiratory behavior is closely monitored in real time under high-density, low-light conditions to accurately identify heat stress and respiratory disorders, factors that directly govern animal welfare and production efficiency in precision livestock farming. This study presents GDLYOLO-Tracker, a deep learning-based framework specifically engineered for nighttime infrared poultry surveillance. To address the unique challenges of nighttime monitoring, the GDLYOLO detector integrates three specialized modules: GhostNet-C2f leverages the inherent feature redundancy in dense flocks to minimize parameter count; Dynamic ATSS facilitates adaptive sample allocation, effectively mitigating precision loss caused by blurred boundaries in low-light and occluded conditions; and LiteShiftHead enhances the detection of fine-grained respiratory behaviors with high computational efficiency. To address the challenges of color information loss and highly similar appearance features in infrared poultry scenes, the system integrates the ByteTrack algorithm. By employing an appearance-free, motion-based association strategy, it ensures accurate identification and consistent tracking even amidst frequent occlusions. Experimental results show that the optimized GDLYOLO model achieves 94.6% recall and 76.1% mAP@0.5-0.95, with a 5.3 MB model size and 7.1 GFLOPs computational cost. Relative to the baseline, this reflects improvements of 1.1% in recall, 1.2% in mAP@0.5-0.95, and a 3.6% reduction in model size. HOTA, MOTA, and IDF1 scores rose by 1.59%, 0.79%, and 2.66%, respectively. Compared with other MOT algorithms, the proposed method yields the highest HOTA, MOTA, and IDF1 scores at 161.11 FPS, demonstrating superior tracking stability and efficiency. Also, the GDLYOLO-Tracker model was applied through a GUI developed using PyQt to support the automatic identification of the abnormal open-mouth behavior (OMB) of poultry. These features enabled stress temperature notifications and proper health assessments to be automated. Field validation at commercial poultry farms confirmed that the system reliably detects OMB and issues prompt alerts via the remote monitoring platform.

Fields of Study & Subjects

Video Surveillance and Tracking Methods Smart Agriculture and AI Advanced Neural Network Applications

Publication details

Journal / source
Smart Agricultural Technology
Volume
14
Pages
102067–102067
Publication date
30 March 2026
Language
EN