CAMLLA-YOLOv8n: Cow Behavior Recognition Based on Improved YOLOv8n

Area of Science:

• Computer Vision and Machine Learning in Agriculture
• Animal Behavior Analysis
• Precision Livestock Farming

Background:

• Monitoring cow behavior is vital for health assessment and farm management.
• Accurate detection of behaviors like grazing, standing, and estrus is challenging under real-world conditions.
• Existing models require optimization for improved performance in agricultural environments.

Purpose of the Study:

• To develop an advanced deep learning model, CAMLLA-YOLOv8n, for accurate Holstein cow behavior recognition.
• To enhance the YOLOv8n architecture with attention mechanisms and improved feature extraction for better detection.
• To validate the model's effectiveness on a comprehensive dataset of Holstein cow behaviors.

Main Methods:

• Proposed CAMLLA-YOLOv8n model integrating Coordinate Attention (C2f-CA) and MLLAttention mechanisms.
• Improved SPPF module (SPPF-GPE) for enhanced small target recognition.
• Utilized Shape-IoU loss for improved bounding box matching and employed hybrid data augmentation.

Main Results:

• CAMLLA-YOLOv8n demonstrated superior Precision compared to YOLOv3-tiny, YOLOv5n/s, YOLOv7-tiny, YOLOv8n/s.
• Achieved significant improvements in Precision (2.18%), Recall (1.62%), mAP@0.5 (1.84%), and mAP@0.5:0.95 (1.77%) over YOLOv8n.
• The model effectively detected Holstein cow behaviors including grazing, standing, lying, licking, estrus, and fighting.

Conclusions:

• The CAMLLA-YOLOv8n model offers accurate and rapid Holstein cow behavior identification in agricultural settings.
• This advancement supports the digitalization and intelligence transformation of animal husbandry.
• Improved cow behavior monitoring can lead to enhanced farm economic benefits and animal welfare.