Camera-based pose estimation is becoming a key sensing approach for automated livestock monitoring, yet the selection of camera geometry, keypoint schema, learning model, and deployment strategy remains highly heterogeneous across species and applications. This systematic review synthesised the literature published or available online between 1 January 2015 and 31 December 2025 to identify task-aligned design patterns, reporting gaps, and practical routes toward on-farm implementation. Searches of ScienceDirect, Scopus, Web of Science, and Embase identified 273 records. After duplicate removal, eligibility screening, and methodological quality appraisal, 114 studies were included in the final synthesis. The reviewed literature shows that cattle and pigs dominate the field, while poultry, horses, goats, sheep, camels, and cross-species studies remain comparatively under-represented. Across applications, side-view RGB is most consistently suited to gait and lameness analysis, whereas fixed top-down RGB offers the best trade-off for group-level behaviour, feeding, tracking, and identity-related tasks. Stereo, RGB-D, and multi-view systems are most justified when the target output depends on absolute geometry, persistent occlusion handling, or three-dimensional posture recovery, particularly in body measurement and weight-estimation pipelines. The synthesis further shows that performance depends more on task-camera-keypoint alignment than on any single model family. However, progress toward robust deployment is limited by inconsistent landmark definitions, fragmented datasets, heterogeneous evaluation metrics, and insufficient reporting of runtime, calibration burden, and farm-system integration. A task-oriented reporting structure and preliminary species-specific core keypoint guidance are therefore proposed to improve comparability and reuse. Overall, livestock pose estimation is sufficiently mature for several structured farm applications, but broader adoption will depend on stronger standardisation, more diverse multi-site datasets, and more explicit validation under real operational conditions.
Menolotto et al. (Thu,) studied this question.