Tracking Ethological Behaviors of Broiler Chickens Through Computer Vision (YoloV11) Using an Android Mobile Device

Ryan Davi Oliveira de Meneses, Glauber da Rocha BalthazararXiv 2026 · DOI · arXiv

This paper presents a practical step toward "Precision Poultry Farming" by bringing state-of-the-art computer vision directly to the farm floor via mobile devices. Researchers developed a YOLOv11-based system to automatically track and classify the behaviors of broiler chickens, providing a scalable tool for real-time animal welfare monitoring.

TL;DR

Researchers successfully deployed a YOLOv11 model on an Android device to classify broiler chicken behaviors (sitting, standing, feeding) with up to 98% accuracy for static postures. This mobile-first approach enables continuous, automated welfare monitoring in commercial poultry houses without requiring expensive server infrastructure.

What Problem Does This Paper Address?

Monitoring the health and welfare of thousands of birds in a climate-controlled poultry house is a monumental task for human workers. Traditional manual observation is subjective, labor-intensive, and intermittent. While computer vision offers a solution, many existing systems rely on high-end servers or cloud processing, which are often impractical for rural farm environments with limited connectivity.

The authors tackle two main challenges:

  1. Automation: Creating a model that can distinguish between subtle ethological behaviors (like the difference between a bird that is just sitting and one that is actively feeding).
  2. Portability: Optimizing these complex deep-learning models to run efficiently on a standard Android smartphone using TensorFlow Lite (TFLite).

Methods

The study utilized the YOLOv11 (You Only Look Once) architecture, specifically the YOLOv11M variant, known for its balance between speed and accuracy. The workflow followed a standard but rigorous pipeline:

  • Training: The model was trained on a workstation (AMD Ryzen 5 5600G / Radeon RX 6600) for 100 epochs.
  • Optimization: To move the model to mobile, the authors converted the trained weights to the TFLite format, employing quantization and compression to ensure the app could utilize native hardware accelerators (DSP/GPU/NPU) on a smartphone.
  • Deployment: A custom Android application was developed using Java and the YOLOv4-TFLite API to process camera frames in real-time.

The model was tasked with identifying three primary classes: sitting, standing, and consuming (feeding or drinking).

Datasets

The researchers used a combination of data sources:

  • Public Repositories: Videos were sourced from platforms like Kaggle and Zenodo, featuring broiler chickens in climate-controlled facilities.
  • Preprocessing: Footage was filtered for quality (minimum 360x480 resolution) and age (birds older than three weeks).
  • Annotation: Video frames were sampled every two seconds and manually labeled. The final training/validation split was 70/30.
  • Unpublished Test Set: For final evaluation, a new dataset of 102 images was used to test the model’s generalization to unseen environments.

Key Insights and Ideas

  • Static vs. Dynamic Bias: The model is significantly better at identifying static postures ("sitting") than dynamic ones ("consuming"). This suggests that while object detection is powerful, capturing the temporal nuances of movement remains a challenge for single-frame analysis.
  • The "Biological Overlap" Problem: A fascinating insight from the confusion matrix analysis is that the model's "errors" often mirror biological reality. Modern broilers, as they gain weight, often feed while partially reclined or static. This creates a visual ambiguity where "sitting" and "feeding" look nearly identical, leading to misclassification that isn't just an algorithmic failure, but a reflection of actual bird behavior in intensive rearing systems.
  • Edge Computing Feasibility: The study proves that the latest YOLO iterations are lightweight enough for modern mobile hardware, achieving high precision without needing a constant link to a powerful backend server.
Figure 5. Confusion Matrix

Key Results

The model demonstrated a mean Average Precision ([email protected]) of 0.417 across all classes during validation. However, performance varied wildly by behavior:

  • Sitting: The star performer, achieving 98.04% accuracy and 100% precision on the test set.
  • Standing: Achieved a solid 83.33% accuracy.
  • Consuming: The most difficult class, with 67.65% accuracy. The primary issue here was false negatives (missing 27% of instances) due to occlusions in dense flocks.
Figure 3. Precision recall curve

Visual validation (Figure 7) confirmed the model's ability to handle high-density populations and varying lighting conditions across different farm installations.

Figure 7. Model validation with different installations

Limitations and Future Work

The authors identify occlusion as a major hurdle; in dense commercial flocks, birds often block the camera's view of one another, leading to missed detections. Furthermore, the low recall for feeding behaviors suggests that single-image detection might need to be supplemented.

Future directions include: * Integrating temporal data (video sequences) to better distinguish dynamic behaviors. * Combining visual data with other sensors, such as thermal imaging or acoustic monitoring (detecting pecking sounds), to resolve the ambiguity between sitting and feeding. * Scaling the system for commercial use to provide automated "inactivity indices" as early warning signs for flock health issues.

Summary updated

Abstract

This study developed and evaluated a computer vision model for automated ethological behavior classification of broiler chickens in climate-controlled poultry houses. Using publicly available videos preprocessed and manually annotated into three classes (sitting, standing, and feeding), the YOLOv11 model was trained and deployed on an Android application via TFLite conversion. Exploratory analysis ensured dataset consistency and visual quality, while evaluation showed class-dependent performance: higher robustness for static postures (sitting, AP=0.531; precision=100%) and lower accuracy for dynamic behaviors (feeding, AP=0.292; recall=0.20). Tests with unseen images demonstrated significant improvements, achieving accuracies of 98.04% for sitting, 83.33% for standing, and 67.65% for feeding. The model showed effective generalization in real farm environments, supporting its potential for continuous behavioral monitoring, early welfare assessments, and decision-making in precision poultry farming. These findings establish a technical foundation for future mobile-based applications at commercial scale, reinforcing the role of computer vision in automated animal welfare evaluation.
BibTeX
@misc{meneses2026tracking,
  author = {Ryan Davi Oliveira de Meneses and Glauber da Rocha Balthazar},
  title = {{Tracking Ethological Behaviors of Broiler Chickens Through Computer Vision (YoloV11) Using an Android Mobile Device}},
  year = {2026},
  eprint = {2611.23007},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.23007},
}
Figure