Computer vision system for assessing pig welfare indicators on carcasses

There is an increasing societal demand for transparency in reporting on the quality of life of farmed animals reared for meat production. High animal welfare standards are also associated with improved efficiency and the sustainability of production systems, but they also play a critical role in ensuring food quality and safety. The routine monitoring of animal welfare is crucial for tracking performance to ensure high welfare standards are met. Traditional on-farm welfare assessments conducted by human observers are subjective and prone to observer bias, time-consuming, costly, and pose risks to biosecurity. Monitoring the welfare of pigs at slaughter provides an option for animal welfare oversight across large numbers of animals to verify and complement on-farm assessments. We present a real-time computer vision system for the automated assessment of pig welfare indicators on carcasses. The system evaluates skin and tail lesions, tail length, and hernias using a modular pipeline that combines YOLOv4 detection, U-Net segmentation, and colorimetric and geometric analysis. The architecture robustness was demonstrated by processing video streams containing specific welfare conditions, yielding accuracies of 93.0% for hernias (16 pigs, 1 min), 86.3% and 90.4% for dorsal and lateral skin lesions (75 pigs, 7 min and 40 pigs, 3 min, respectively), and 86.8% for tail lesions (63 pigs, 5 min). Tail length is estimated via a custom segmentation and curve-fitting process, with a root mean squared error (RMSE) of 4.45 cm. Operating at 30.31 FPS, the framework offers a scalable and objective solution for real-time welfare monitoring in industrial settings.