SDGTrack: A Multi-Target Tracking Method for Pigs in Multiple Farming Scenarios.

Area of Science:

• Agricultural technology
• Computer vision
• Animal science

Background:

• Multi-object tracking (MOT) algorithms are crucial for pig identification and health monitoring in modern farming.
• Existing MOT models face challenges generalizing to diverse pig farming environments due to significant variations.
• Lack of adaptability limits the efficiency and intelligence of current pig management systems.

Purpose of the Study:

• To develop an advanced MOT method, SDGTrack, that improves generalization across varied pig farming scenarios.
• To enhance the adaptability of tracking models to different environmental domains within pig farming.
• To provide a robust technical foundation for intelligent pig monitoring in diverse settings.

Main Methods:

• Proposed the SDGTrack method, integrating domain adaptability and an optimized tracking strategy.
• Constructed a comprehensive multi-scenario dataset using public and private data from ten distinct pig farming environments.
• Evaluated performance using daytime scenes for training and remaining daytime/nighttime scenes for validation.

Main Results:

• SDGTrack achieved a Multi-Object Tracking Accuracy (MOTA) of 80.9% and an IDF1 score of 85.1%.
• The method significantly reduced ID switches (IDSW) by 94.6% compared to the original CSTrack.
• Demonstrated substantial improvements over CSTrack, with MOTA increasing by 16.7% and IDF1 by 33.3%.

Conclusions:

• SDGTrack exhibits robust tracking capabilities in novel and diverse pig farming environments.
• The method significantly enhances the generalization of group pig tracking technology.
• Provides a strong foundation for intelligent pig monitoring systems across different farming settings.