Deep learning pose estimation for multi-cattle lameness detection.

Area of Science:

• Agricultural Engineering
• Computer Science
• Animal Science

Background:

• Lameness is a significant welfare and economic issue in dairy cattle.
• Accurate, real-time lameness detection is crucial for timely intervention and herd management.
• Existing methods may lack automation, scalability, or precision.

Purpose of the Study:

• To develop a fully automated, deep learning-based system for real-time lameness detection in multiple cows.
• To utilize computer vision and pose estimation for accurate analysis of cow posture and gait.
• To create a system deployable on commercial dairy farms.

Main Methods:

• Employed a modified Mask R-CNN for cow pose estimation, identifying key points for back arching and head position.
• Utilized the SORT algorithm for real-time tracking of individual cows within video sequences.
• Integrated features using the CatBoost gradient boosting algorithm for lameness classification.
• Validated the system against ground truth data from accredited mobility scorers.

Main Results:

• Achieved 100% accuracy in threefold lameness detection and 94% accuracy in lameness severity classification.
• Demonstrated high precision (Cohen's kappa = 0.8782, precision = 0.8650, recall = 0.9209).
• The system accurately analyzed posture and gait simultaneously with 94-100% accuracy.

Conclusions:

• The developed deep learning system offers a highly accurate and automated solution for real-time lameness detection in dairy cows.
• The system's ability to track and analyze multiple cows simultaneously enhances its practical applicability in farm settings.
• This technology has the potential to significantly improve animal welfare and farm economics through early lameness intervention.