Answers are grounded in this paper and cite their sources.
Automated behavior monitoring is essential for advancing precision livestock farming, improving animal welfare, and enabling early disease detection. This study systematically evaluated six lightweight object detection architectures, YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLOv11, and YOLOv12, for detecting pig postural (Lateral Lying, Sternal Lying, Standing, and Sitting) and nutritive (Feeding, Drinking, and Not Feeding/Drinking) behaviors. Three detection models were developed, and Grad-CAM was applied to visualize the key visual features underlying model predictions. Results showed that task formulation had a greater impact on performance than architectural differences. Nutritive models achieved the highest detection performance (average mAP50 = 95.8–96.6%), followed by posture models (89.9–91.1%), while combined models were the most challenging (79.8–81.9%) due to increased classification complexity and potentially due to class imbalance. Detection difficulty was behavior-specific: Standing, Feeding, and Not Feeding/Drinking were identified most reliably, whereas Sitting, Drinking, Sitting Drinking, and Sternal Lying Feeding remained challenging. Grad-CAM analysis confirmed that model predictions were based on biologically meaningful features, integrating global body posture with localized head–resource interactions. Based on the present study dataset, YOLOv9 provided the highest overall detection performance, achieving the highest mAP50 and mAP50-95 in posture and combined models while matching YOLOv12 in the nutritive model. However, it required the longest training time and the slowest inference. YOLOv8 achieved the lowest overall latency, whereas YOLOv10 provided a balanced trade-off between accuracy and speed. These findings indicate that future progress in pig behavior recognition will depend not only on livestock-specific architectural innovations and more representative benchmark datasets.
No takes yet. Share an insight, caveat, or question.
Lagua et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: