Study on Poultry Pose Estimation Based on Multi-Parts Detection.

Area of Science:

• Computer Vision
• Animal Science
• Machine Learning

Background:

• Accurate poultry pose estimation is vital for disease prediction and managing poultry health.
• Automatic pose recognition in group-fed chickens presents challenges for large-scale farm monitoring.

Purpose of the Study:

• To develop and evaluate a deep neural network (DNN) technique for single broiler chicken pose estimation.
• To compare the DNN method's performance against established algorithms like SSD, YOLOV3, RetinaNet, and Faster R-CNN.

Main Methods:

• Utilized computer vision technology and a deep neural network (DNN) for pose estimation.
• Compared the proposed DNN method with Single Shot MultiBox Detector (SSD), You Only Look Once (YOLOV3), RetinaNet, and Faster R-CNN algorithms.

Main Results:

• The proposed DNN method achieved a precision standard deviation of 0.0128 with 95% confidence (0.9218 ± 0.0048).
• The method demonstrated a recall standard deviation of 0.0266 with 95% confidence (0.8996 ± 0.0099).

Conclusions:

• Successful broiler chicken pose estimation facilitates the detection of abnormal poultry behavior.
• The developed method shows promise for improving poultry health monitoring and can be further refined.