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

Area of Science:

• Animal Science
• Computer Vision
• Machine Learning

Background:

• Accurate dairy cow behavior recognition is crucial for smart livestock farming, impacting health monitoring, welfare assessment, and early warning systems.
• Recognizing fine-grained behaviors like feeding, drinking, and rumination is challenging in real barn environments due to occlusion, complex backgrounds, subtle motion, and class imbalance.

Purpose of the Study:

• To develop an advanced dairy cow behavior recognition method integrating temporal modeling and behavioral priors.
• To improve the accuracy and reliability of non-contact monitoring of key dairy cow behaviors in practical barn management.

Main Methods:

• Utilized the Contrastive Language-Image Pre-training (CLIP) visual encoder for feature extraction.
• Integrated two temporal adapters to model dynamic information across video frames.
• Decoupled behavior recognition into posture and action recognition, incorporating a behavioral prior loss to constrain improbable posture-action combinations.

Main Results:

• Achieved a five-class accuracy of 75.45% and a five-class Macro-F1 score of 0.7246 on the test set.
• Obtained an Action Macro-F1 score of 0.7605.
• Outperformed the baseline CLIP model and several other representative video recognition models in dairy cow behavior recognition tasks.

Conclusions:

• The proposed method effectively integrates temporal modeling and behavioral priors for enhanced dairy cow behavior recognition.
• Demonstrated superior performance compared to existing models, indicating its potential for practical barn management and non-contact monitoring.
• Supports improved animal welfare and health management through accurate, automated behavior analysis.