Deep learning pose detection model for sow locomotion.

Area of Science:

• Animal Science
• Computer Vision
• Machine Learning

Background:

• Lameness in sows causes pain and welfare issues, often going undetected in early stages.
• Automated, non-invasive systems are needed for precise and reliable lameness detection.
• Current methods lack objectivity and precision, hindering timely intervention.

Purpose of the Study:

• To create a comprehensive image and video repository of sows with varying locomotion scores.
• To develop a computer vision model for automatic identification and tracking of sow body keypoints.
• To facilitate kinematic studies for objective lameness detection using deep learning.

Main Methods:

• Collected 2D videos of sows with different lameness scores on a specialized farm setup.
• Utilized two stereo cameras for video recording.
• Annotated videos using the Zinpro Locomotion Score System by 13 locomotion experts.
• Trained and tested deep learning models using the SLEAP framework.

Main Results:

• Developed models accurately tracked 6 (lateral) and 10 (dorsal) skeleton keypoints.
• Achieved high performance metrics: average precision (0.90 lateral, 0.72 dorsal), low pixel distance (6.83 lateral, 11.37 dorsal), and high similarity (0.94 lateral, 0.86 dorsal).
• Demonstrated the potential for objective posture estimation and lameness scoring.

Conclusions:

• The developed computational models serve as a Precision Livestock Farming tool for automatic pig posture analysis.
• The annotated video repository is valuable for teaching and research in animal locomotion.
• An automated system based on these findings can objectively assess sow locomotion scores, enhancing animal welfare.