Monitoring dairy cow behavior is essential for health management and overall welfare. Behaviors such as standing, lying, eating, and drinking provide key indicators of well-being—prolonged lying can signal hoof disease, while reduced food intake may indicate digestive issues. Traditionally reliant on visual inspections, these methods can be labor-intensive and disruptive to animals.
This study introduces Res-DenseYOLO, an advanced computer vision model tailored for detecting cow behaviors in real-time. By incorporating innovations such as a dense module for enhanced feature extraction, a CoordAtt attention mechanism, and a SioU loss function, the model achieves significant improvements in precision, recall, and mean average precision (mAP) compared to existing methods. Tested on over 5500 images from dairy farm video footage, Res-DenseYOLO achieved a precision of 94.7%, recall of 91.2%, and mAP of 96.3%, outperforming state-of-the-art models like YOLOv4 and Fast-RCNN.
This research demonstrates the model’s potential for improving animal welfare and production efficiency, offering a non-invasive and scalable solution for real-world dairy farm applications.
