This study used video data from 10 Holstein cows before calving. Three key prepartum behaviors were targeted:
- Tail Raising: the cow lifts its tail for more than 3 seconds.
- Head to Abdomen: the cow turns its head toward its abdomen, often in response to discomfort.
- Lying-Standing Transition: the cow changes position between lying and standing.
Fifteen body parts were labeled on each cow. Using DeepLabCut, a deep learning tool for pose estimation, these keypoints were tracked across video frames. Then, random forest classifiers were trained to detect the three behaviors automatically.
To explore how these behaviors changed in the 24 hours before calving, zero-inflated negative binomial (ZI-NB) models were applied. Finally, a prediction model was developed to classify whether a cow was within 5 hours of calving, with Tail Raising showing the strongest predictive value.
