Pose Estimation and Behavior Classification of Jinling White Duck Based on Improved HRNet

Area of Science:

• Computer Vision
• Animal Science
• Artificial Intelligence

Background:

• Accurate pose information is crucial for assessing duck physiological health, welfare, and environmental comfort in breeding.
• Existing methods may lack the precision or adaptability required for comprehensive duck pose analysis.

Purpose of the Study:

• To develop an automatic and accurate multi-pose estimation method for ducks.
• To enhance the detection of keypoints across various duck behaviors for improved monitoring.

Main Methods:

• Proposed a novel pose estimation model by integrating HRNet-32 with Convolutional Block Attention Modules (CBAM).
• Optimized the HRNet-32 architecture and embedded multiple CBAM modules for enhanced feature extraction.
• Evaluated model performance on Cherry Valley ducklings of different ages under varying illumination conditions.

Main Results:

• The HRNet-32-CBAM model achieved an average precision (AP) of 0.943 on the duck pose estimation dataset.
• Demonstrated strong generalization ability across different ages, breeds, and farming modes.
• Confirmed real-time performance for multi-pose estimation on various image resolutions.

Conclusions:

• The developed HRNet-32-CBAM model offers an accurate and efficient solution for duck multi-pose estimation.
• This method provides a valuable technical reference for intelligent poultry farming and animal welfare monitoring.
• The study highlights the potential of deep learning for advancing precision livestock management.