Answers are grounded in this paper and cite their sources.
Behavioural monitoring is central to understanding how animals respond to naturally occurring and human‐induced environmental change, yet scalable approaches for behavioural data collection, analysis and interpretation remain limited. Zoological institutions provide a valuable setting for developing and validating automated behavioural monitoring tools, as they encompass complex environments containing wildlife while offering a degree of control relative to conditions in the wild. Although recent advances in machine learning (ML) facilitate behavioural inference from video data, few approaches translate these capabilities into real‐time, alert‐enabled monitoring systems capable of continuous deployment. We present ROAM (real‐time objective animal monitoring), a generalizable ML framework that integrates computer vision‐based models in an end‐to‐end pipeline to support automated animal behaviour surveillance in both wild and managed care settings, using polar bears ( Ursus maritimus ), giraffes ( Giraffa reticulata ), zebras ( Equus spp.) and Siberian jays ( Perisoreus infaustus ) as focal species. This framework conducts real‐time classification of routine behaviours (e.g. foraging, locomotion, resting, swimming and head swinging), while repetitive behaviours such as loop swimming are inferred using logic‐based heuristics applied to frame‐wise model outputs. The system supports customizable alert thresholds based on behaviour duration and/or frequency, with automated notifications delivered to designated personnel. The ROAM toolkit was applied to publicly available datasets spanning distinct taxa and study contexts, where it generated behavioural alerts and automated post‐monitoring summaries with modification only to the detection and behaviour‐definition layers, requiring minimal adjustment to the core pipeline. Beyond real‐time alerts, ROAM streamlines downstream data processing and generates interpretable outputs, including activity budgets and spatiotemporal visualizations. Altogether, this open‐source pipeline provides a transferable and cost‐efficient method for converting continuous video streams into structured behavioural datasets suitable for quantitative analysis across welfare science, applied behavioural research and conservation management.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: