Leveraging keypoints for multi-camera tracking of Jersey cattle in calving pens

Accurate multi-cameratracking of individuals is essential to advance precision livestock managementand automated behaviour monitoring, yet reliably matching indistinguishableanimals such as Jersey cows across different viewpoints remains a challenge. Re-identificationapproaches often rely on visually distinctive features that Jersey cattle lack,leading to frequent ID mismatches and reduced tracking accuracy. Wideside-angled views introduce further challenges such as occlusion, scalevariation, and perspective distortion. This study developed a multi-cameratracking framework capable of tracking nine cows across two cameras covering a108 m² calving pen. We trained a multi-class YOLOv11-Pose model on 1035 images,distinct from the tracking dataset, that can extract four dorsal keypoints aswell as classify standing and lying postures with a 98.2% accuracy. Weextracted and annotated 1000 frames per camera for our tracking dataset. Weutilised an improved ByteTrack, with an additional matching step incorporatingkeypoints, over a sliding window to yield tracklets per camera. Forcross-camera matching, we computed a cost matrix by measuring the reprojectionerror from triangulated 3D keypoints back onto each 2D camera plane. TheHungarian algorithm matched tracklets across the two cameras, assigning eachcow a consistent global ID regardless of viewpoint. Our framework achievedmulti-object tracking accuracies (MOTA) of 92.43% in Camera 1 and 97.02% inCamera 2, illustrating reliable tracking for both cameras. This studyunderscores the viability of a multi-camera, keypoint-driven approach fortracking visually similar animals, which can be extended to larger-scale farmswhere minimising camera infrastructure is a critical necessity.