Beyond proximity: Pose-based interaction classification for dairy cattle social behavior analysis

Precision livestock farming requires objective, scalable assessment of social behavior, yet most automated approaches infer interactions using static proximity thresholds that fail to distinguish affiliative from agonistic behaviors in complex barn environments. This limitation restricts the interpretability and practical utility of vision-based social network analysis. A pose-driven approach is introduced in which interaction valence is inferred from the spatiotemporal dynamics of anatomical keypoints rather than spatial co-location alone. By encoding motion signatures from keypoint trajectories, the method captures geometric and temporal features that differentiate social behaviors without reliance on pixel-level appearance. The system is implemented as an end-to-end pipeline integrating YOLOv11 for detection (mAP@0.50: 96.24%), supervised identification (98.24%), ByteTrack for tracking (81.96%), ZebraPose for 27-point keypoint estimation, and a two-stage temporal convolutional network for interaction classification. On a held-out set of annotated interaction clips from a commercial dairy barn, the classifier achieved 86.23% accuracy, outperforming proximity-based baselines by 25 percentage points. Validation is conducted on a single-site, single-day dataset (6 cows), providing controlled evidence of feasibility rather than population-level generalization. This work establishes a controlled proof-of-concept for pose-based interaction inference, demonstrating that behavioral valence can be derived from keypoint dynamics, while large-scale generalization across herds, environments, and seasons remains a subject of ongoing investigation.