Automated 3D sheep posture estimation and telemetric pressure synchronization via a markerless computer vision pipeline

Body posture is a key determinant of multi-compartmental pressure dynamics, yet continuous posture monitoring in chronic large animal studies remains technically challenging. To investigate physiological pressure dynamics as a function of body posture in an ovine model, we present a markerless computer vision pipeline that calculates three-dimensional (3D) pose estimates of an individual sheep during training sessions and automatically synchronizes posture with telemetric pressure recordings. First, two-dimensional (2D) joint localization was performed using a combination of Faster R-CNN and Lite-HRNet, achieving a mean average precision (mAP) of 77.4%. Second, we triangulated the detected 2D keypoints into 3D space and leveraged a skeletal and temporal prior to mitigate 2D localization errors. Third, we computed anatomical angles and used a multilayer perceptron to classify five structured training postures (standing, walking up a bridge, walking down a bridge, kneeling, sitting) achieving an overall accuracy of 95.2% on a held-out test set of 647 frames drawn from six sheep, and a leave-one-subject-out cross-validated accuracy of 68.8% when tested on animals unseen during training. Finally, the telemetric recordings of physiological pressures are synchronized to the respective posture allowing for retrospective analysis of body posture in relation to pressure dynamics. This pipeline lays the foundation for future investigations into posture-dependent pressure recordings, with implications for both animal modeling and clinical translation to humans.