Broiler Mobility Assessment via a Semi-Supervised Deep Learning Model and Neo-Deep Sort Algorithm.

Area of Science:

• Animal Science
• Computer Science
• Agricultural Engineering

Background:

• Broiler mobility is a key welfare indicator, but traditional assessment is labor-intensive and lacks real-time insights.
• Automated methods are needed for efficient and accurate monitoring of broiler welfare.
• Deep learning offers potential for advanced animal behavior analysis.

Purpose of the Study:

• To develop and validate a semi-supervised deep learning model for individual broiler mobility tracking.
• To enhance the accuracy and efficiency of broiler welfare assessment through automated mobility analysis.
• To quantify individual and flock-level broiler mobility, addressing challenges like occlusion and detection errors.

Main Methods:

• A semi-supervised deep learning (DL) approach using YOLOv5 (You Only Look Once version 5) was employed.
• The YOLOv5 model was initially trained on labeled images and then fine-tuned using pseudo-labeling via semi-supervised learning (SSL).
• A neo-deep sort algorithm, combined with the trained YOLOv5, was used for broiler detection, tracking, and mobility quantification.

Main Results:

• Semi-supervised learning significantly improved the YOLOv5 model's detection mean average precision (mAP) from 81% to 98%.
• The combined model achieved approximately 80% validation accuracy in estimating individual broiler hourly distances compared to manual measurements.
• The system successfully tracked 28 broilers, quantifying hourly and daily travel distances and speeds while mitigating detection issues.

Conclusions:

• The proposed semi-supervised DL model provides an accurate and efficient tool for automated broiler mobility tracking.
• This technology enables precise quantification of individual and flock-level broiler movement, crucial for welfare assessment.
• The method overcomes common challenges in animal tracking, offering a robust solution for precision livestock farming.