FSCA-YOLO: An Enhanced YOLO-Based Model for Multi-Target Dairy Cow Behavior Recognition

Area of Science:

• Computer Vision
• Animal Science
• Machine Learning

Background:

• Object recognition models in dairy farming face challenges with complex backgrounds and cow occlusions, leading to detection errors.
• Accurate cow behavior recognition is crucial for optimizing dairy farm management and animal welfare.

Purpose of the Study:

• To develop an improved multi-object cow behavior recognition model, FSCA-YOLO, addressing limitations of existing systems in real-world dairy environments.
• To enhance the accuracy and efficiency of cow behavior recognition for practical applications in livestock monitoring.

Main Methods:

• An improved YOLOv11 framework was utilized, incorporating the FEM-SCAM module with CoordAtt for feature focus and a small object detection head for distant targets.
• The SIoU loss function replaced the original loss function to boost recognition accuracy and convergence speed.
• OpenCV was integrated for specific behavior recognition and in-region counting functionalities.

Main Results:

• FSCA-YOLO achieved superior performance over baseline YOLOv11, with precision at 95.7%, recall at 92.1%, and mean average precision (mAP) at 94.5%.
• The model demonstrated significant improvements of 1.6% in precision, 1.8% in recall, and 2.1% in mAP compared to the baseline.
• The enhanced model accurately extracts cow features in complex farming settings, proving its practical utility.

Conclusions:

• FSCA-YOLO provides a robust and reliable vision-based solution for cow behavior recognition in challenging dairy farming conditions.
• The model's integration with OpenCV enhances its adaptability for diverse behavior identification and counting needs in multi-cow systems.
• This research contributes to advancing automated livestock monitoring through improved object detection and behavior analysis.