Deep Learning Methods for Tracking the Locomotion of Individual Chickens.

Area of Science:

• Animal Science
• Computer Vision
• Agricultural Technology

Background:

• Poultry locomotion is crucial for assessing animal health, welfare, and productivity.
• Traditional methods for monitoring poultry movement are often intrusive and can alter behavior.
• There is a need for non-invasive, accurate methods to analyze poultry locomotion.

Purpose of the Study:

• To adapt and evaluate an enhanced Track Anything Model (TAM) for non-intrusive poultry locomotion analysis.
• To assess TAM's performance in tracking and analyzing the movement of various chicken types.
• To compare TAM's effectiveness against existing state-of-the-art models like YOLOv5 and YOLOv8.

Main Methods:

• Utilized an enhanced Track Anything Model (TAM) for tracking chickens in diverse experimental settings.
• Employed a dataset including dyed and undyed broilers and layers for model adaptation and evaluation.
• Quantified tracking performance using intersection over union (mIoU) and speed accuracy with root mean square error (RMSE).

Main Results:

• The enhanced TAM demonstrated superior segmentation and tracking capabilities for poultry.
• Achieved high mIoU values (93.12%) across different chicken categories, outperforming other models.
• Showcased accurate speed detection with an RMSE of 0.02 m/s, validating its precision.

Conclusions:

• The enhanced TAM provides a technologically advanced, consistent, and non-intrusive method for poultry locomotion analysis.
• TAM is a potent tool for detailed poultry behavior monitoring, contributing to improved animal welfare.
• The model's potential extends to broader livestock monitoring, enhancing farm management practices.