Bovine-LTID: A vision-based spatio-temporal identity system for multi-camera cattle tracking and long-term behavior monitoring

Multi-camera video monitoring of group-housed cattle is essential for health assessment in precision livestock farming, offering a scalable, non-invasive alternative to labor-intensive manual observations and restrictive wearable sensors. However, existing systems for cross-camera identity matching typically rely on computationally intensive appearance-based Re-identification (ReID). These approaches often struggle in dense barns due to frequent mutual occlusions, viewpoint variations, and the high visual similarity of cattle coat patterns. To address these limitations, we propose Bovine Long-Term Identity (Bovine-LTID), a resource-efficient, ReID-embedding-free spatio-temporal tracking framework for multi-camera behavior monitoring. Instead of relying on ambiguous appearance features, Bovine-LTID maintains identity consistency through spatial geometry and temporal logic. The system integrates a behavior-aware detector (YOLOv9s) with a cattle-optimized single-view tracker (CowTracker) to recover identities during prolonged occlusions. For cross-view synchronization, trajectories are projected into a unified ground-plane coordinate system via homography calibration and managed by a centralized Global ID Pool. Trained on NVIDIA H200 GPU, the detection module achieves a test-set mAP50 of 0.944, while the complete pipeline operates at approximately 13 FPS on consumer-grade hardware (NVIDIA RTX 4060 Laptop, 8 GB VRAM). The cross-camera fusion mechanism reduces global raw identity switches by 87.6% over single-camera baselines and achieves a 99.33% cross-view identity consistency. These findings present a practical methodology for scalable, automated herd welfare assessment spanning from high-performance training to edge deployment.