Enhanced Swine Behavior Detection with YOLOs and a Mixed Efficient Layer Aggregation Network in Real Time

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Effective livestock management is crucial due to an aging agricultural workforce and large-scale farming.
• Manual observation methods for monitoring livestock behavior are becoming insufficient.
• Automated solutions are needed for real-time monitoring in modern agriculture.

Purpose of the Study:

• To develop an innovative system for real-time sow and piglet behavior detection.
• To enhance feature learning capabilities in animal behavior analysis.
• To improve the efficiency and reliability of livestock monitoring systems.

Main Methods:

• Developed a mixed-ELAN system, replacing standard convolutions with MixConv of diverse kernel sizes.
• Applied the enhanced architecture to YOLOv7 and YOLOv9 object detection models.
• Evaluated performance using k-fold cross-validation (k=3) for reliability.

Main Results:

• The mixed-ELAN architecture improved YOLOv7 and YOLOv9 performance by 1.5% and 2%, respectively.
• Achieved mean average precision scores of 0.805 (YOLOv7) and 0.796 (YOLOv9).
• Significantly enhanced detection of critical piglet behaviors like crushing and lying down.

Conclusions:

• The mixed-ELAN system demonstrates significant improvements in real-time livestock behavior detection.
• AI and computer vision offer substantial benefits for animal welfare and farm management efficiency.
• This research sets new benchmarks for smart farming technologies and livestock management innovation.