A Clip-Based Dairy Cow Behavior Recognition Method Integrating Temporal Modeling and Behavioral Priors

Accurate dairy cow behavior recognition is important for health monitoring, welfare assessment, and early warning in smart livestock farming. However, recognizing fine-grained behaviors such as feeding, drinking, and rumination remains difficult in real barns because of occlusion, complex backgrounds, subtle motion changes, and class imbalance. This study proposes a behavior recognition method that integrates temporal modeling and behavioral priors. The Contrastive Language-Image Pre-training (CLIP) visual encoder is used as the feature extraction backbone, while two temporal adapters are introduced to model dynamic information across consecutive video frames. Dairy cow behavior recognition is further decoupled into posture recognition and action recognition, and a behavioral prior loss is designed to softly constrain unlikely posture-action combinations, such as lying with feeding or lying with drinking. On the test set, the proposed method achieves a five-class accuracy of 75.45%, a five-class Macro-F1 of 0.7246, and an Action Macro-F1 of 0.7605, outperforming the CLIP baseline and several representative video recognition models. These results indicate that the proposed method can support non-contact monitoring of key dairy cow behaviors for practical barn management.