Deep Learning-Based Multi-Animal Tracking for Behavioral Analysis: From Identity Preservation and Pose Estimation to Social Interaction Inference

Deep learning-based video analysis has enabled markerless pose estimation, identity-preserved tracking, and quantitative behavioral analysis in a range of experimental animal systems. However, extending these approaches from single animals to interacting groups introduces persistent challenges, including body-part assignment, identity preservation, occlusion, crowding, missing keypoints, and reduced reliability during close social interactions. This structured narrative review examines multi-animal tracking based primarily on 2D RGB video and its connection to downstream behavioral analysis. Literature was identified by searching PubMed, bioRxiv, arXiv, Scopus, and IEEE Xplore on 15 September 2026, supplemented by citation tracking, and incorporated iteratively according to relevance to the review topics. The review highlights that tracking accuracy alone does not determine whether data are suitable for a behavioral question. Identity errors, missing keypoints, and temporal processing can affect behavioral classification, social metrics, and directed interaction estimates. It also emphasizes that the required tracking representation depends on the intended analysis: identity-preserved trajectories may support movement and group-level measures, whereas full-body pose may be required for posture- or orientation-dependent interactions. Reliable use of multi-animal tracking requires quality control and data preparation matched to the intended behavioral analysis.