[Paper Review] Broiler-Net: A Deep Convolutional Framework for Broiler Behavior Analysis in Poultry Houses

Tahereh Zarrat Ehsan, Seyed Mehdi Mohtavipour|arXiv (Cornell University)|Jan 22, 2024

Animal Behavior and Welfare Studies4 citations

TL;DR

Broiler-Net is a real-time deep learning framework for detecting abnormal broiler chicken behaviors—specifically inactivity and huddling—in cage-free poultry houses. It integrates YOLO-based detection, a fast tracking module, and behavior classification to enable timely interventions, achieving high accuracy in real-world video streams with strong performance on benchmark datasets.

ABSTRACT

Detecting anomalies in poultry houses is crucial for maintaining optimal chicken health conditions, minimizing economic losses and bolstering profitability. This paper presents a novel real-time framework for analyzing chicken behavior in cage-free poultry houses to detect abnormal behaviors. Specifically, two significant abnormalities, namely inactive broiler and huddling behavior, are investigated in this study. The proposed framework comprises three key steps: (1) chicken detection utilizing a state-of-the-art deep learning model, (2) tracking individual chickens across consecutive frames with a fast tracker module, and (3) detecting abnormal behaviors within the video stream. Experimental studies are conducted to evaluate the efficacy of the proposed algorithm in accurately assessing chicken behavior. The results illustrate that our framework provides a precise and efficient solution for real-time anomaly detection, facilitating timely interventions to maintain chicken health and enhance overall productivity on poultry farms. Github: https://github.com/TaherehZarratEhsan/Chicken-Behavior-Analysis

Motivation & Objective

• To develop a real-time, accurate system for detecting abnormal behaviors in broiler chickens within cage-free poultry houses.
• To address critical challenges in poultry health monitoring, including early detection of inactivity and huddling, which signal potential disease or stress.
• To improve farm productivity and reduce economic losses through timely, data-driven interventions.

Proposed method

• Utilizes a state-of-the-art YOLO-based object detection model to identify individual broilers in video frames.
• Employs a fast, lightweight tracking module to maintain identity continuity across consecutive video frames.
• Applies a behavior classification head to detect abnormal patterns—specifically inactivity and huddling—based on spatial and temporal features.
• Processes video streams in real time to enable immediate alerts for farm managers.
• Trained and evaluated on a custom dataset collected from real poultry house environments.
• Integrates detection, tracking, and behavior analysis into a unified, end-to-end deep learning pipeline.

Experimental results

Research questions

• RQ1 Can a deep learning framework accurately detect inactivity and huddling behaviors in broiler chickens in real time within cage-free poultry houses?
• RQ2 How does the integration of object detection and multi-object tracking improve behavior analysis accuracy compared to standalone methods?
• RQ3 To what extent does the proposed framework support timely intervention for maintaining flock health and productivity?

Key findings

• The framework achieves high detection accuracy for both inactivity and huddling behaviors, demonstrating robustness in real-world poultry house conditions.
• Real-time inference is feasible, enabling immediate detection and potential alerts for farm operators.
• The integration of YOLO-based detection with a fast tracker significantly improves tracking stability and behavior classification performance.
• The system outperforms baseline models in both detection speed and accuracy on the custom dataset.
• Quantitative results show strong mAP and F1-score performance, indicating reliable behavior recognition.
• The framework is deployable in practical farm settings due to its efficiency and real-time processing capability.