MCP: Multi-Chicken Pose Estimation Based on Transfer Learning.

Area of Science:

• Computer Vision
• Animal Behavior
• Machine Learning

Background:

• Accurate poultry behavior analysis is crucial for effective farm management.
• Pose estimation is a key component of behavior analysis, but multi-chicken scenarios are challenging.
• Existing methods lack specialized solutions for estimating the postures of multiple chickens simultaneously.

Purpose of the Study:

• To develop and evaluate a deep learning-based system for multi-chicken pose estimation (MCP).
• To accurately identify the position and pose of individual chickens within an image.
• To advance poultry behavior analysis through improved pose estimation techniques.

Main Methods:

• A top-down pose estimation approach was employed using deep learning.
• A chicken detector identified individual chickens, followed by a pose estimation network utilizing transfer learning.
• Performance was evaluated using metrics like mean average precision (mAP), mean average recall (mAR), percentage of correct keypoints (PCKs), and root mean square error (RMSE).

Main Results:

• The MCP system achieved a mean average precision (mAP) of 0.652 and a mean average recall (mAR) of 0.742.
• The percentage of correct keypoints (PCKs) reached 0.789, with a root mean square error (RMSE) of 17.30 pixels.
• This research marks the first application of transfer learning for multi-chicken pose estimation.

Conclusions:

• The proposed multi-chicken pose (MCP) system demonstrates effective deep learning-based pose estimation for multiple chickens.
• The method provides a significant advancement for poultry behavior analysis and related research.
• This approach offers a new direction for future studies in automated poultry monitoring and welfare assessment.