Toward intelligent pig farming: Advancing health and welfare monitoring through computer vision–based behavioral analysis

Modern pig production is increasingly moving toward smart and intelligent farming systems, integrating advanced technologies such as artificial intelligence for precision management. Camera-based computer vision has emerged as a promising non-contact sensing tool for monitoring pig health and welfare through behavioral analysis, given the strong association between behavioral patterns and physiological status. Following a PRISMA-based search, 45 studies were included for critical synthesis on computer vision-based behavioral analysis in pigs, with emphasis on detectable behavioral indicators, imaging configurations, model architectures, post-processing strategies, and deployment constraints. Findings show that posture and nutritive behaviors are the most mature targets for monitoring because they are visually distinctive, biologically informative, and compatible with lightweight object-detection pipelines. In contrast, aggression, biting, enrichment interaction, suckling, and individual activity metrics require stronger temporal modeling, segmentation, or multi-object tracking, which increases annotation burden, computational cost, and susceptibility to occlusion and identity errors. Recent pig-specific modifications to YOLO, transformer-inspired, segmentation, and tracking models have improved performance through attention mechanisms, multi-scale feature fusion, occlusion-aware regression, lightweight convolution, and detector-tracker coupling. However, high reported accuracies often reflect controlled datasets rather than commercial robustness. Critical barriers include limited multi-farm validation, inconsistent annotation and post-processing rules, insufficient anomaly-detection thresholds, weak long-term identity preservation, and hardware constraints for edge deployment. Future efforts should prioritize developing standardized benchmarks, biologically grounded behavioral metrics, explainable and domain-adaptive artificial intelligence, and edge-ready architectures that convert model outputs into interpretable management alerts. This review provides a deployment-oriented framework for advancing computer vision from experimental behavior recognition toward scalable behavioral monitoring and intelligent decision support in pig production systems.