Long-Term identity-consistent multi-pig tracking in group-housed pens

In intelligent and precision livestock farming, continuous individual-level tracking of group-housed pigs is essential for behavior monitoring and health management. However, in real-world group-housing environments, high appearance similarity among pigs, frequent occlusions, intensive interactions, and irregular motion patterns cause trajectory fragmentation and identity drift. Therefore, a long-term multi-pig tracking framework tailored for group-housed pig barn scenarios is developed in this study. Within a standard “detection–association–trajectory management” pipeline, data association is performed by jointly exploiting appearance features and motion information, with Re-identification-based appearance modeling. In addition, an identity memory bank and a trajectory relinking mechanism are introduced to enhance identity recovery following occlusions and reduce identity switches (IDSWs). The proposed framework was evaluated on real-world pig barn surveillance video data. Under same-scene conditions, the model demonstrated improved identity consistency and high tracking accuracy, achieving a multiple object tracking accuracy (MOTA) of 0.9010, an identity F1 Score (IDF1) of 0.6762, and an IDSW of 24, outperforming BoT-SORT (MOTA = 0.8673, IDF1 = 0.5675, IDSW = 40), ByteTrack (MOTA = 0.8585, IDF1 = 0.5735, IDSW = 30) and DeepSORT (MOTA = 0.8640, IDF1 = 0.5615, IDSW = 40). Moreover, under more challenging cross-scene conditions, partial generalization was still observed (MOTA = 0.7011, IDF1 = 0.0950, IDSW = 405), indicating transferability to scene variations, severe occlusions, and imaging noise. Overall, the proposed framework provides more reliable long-term individual-level trajectory acquisition for group-housed pigs, supporting behavior analysis and management decision-making in intelligent livestock farming. The method has been made publicly available on GitHub (https://github.com/glimmerc33-ui/pig_track.git). © 2026 Elsevier B.V.