Multicow pose estimation based on keypoint extraction.

Area of Science:

• Precision Livestock Farming
• Computer Vision
• Animal Welfare Monitoring

Background:

• Automatic monitoring of dairy cows is crucial for assessing their health and well-being in precision farming systems.
• Estimating cow poses accurately is challenging due to visual similarities between animals.

Purpose of the Study:

• To develop and validate a multicow pose estimation algorithm for real-time monitoring of dairy cows in farm environments.
• To assess the algorithm's performance in detecting multiple cows, estimating keypoints, and classifying typical poses.

Main Methods:

• Established a monitoring system using 175 surveillance videos to create datasets for object detection and pose estimation.
• Fine-tuned the You Only Look Once (YOLO)v4 model for multi-cow detection and employed a 2D pose detection model for keypoint and Part Affinity Field extraction.
• Developed a pose classification network to identify standing, walking, and lying postures based on extracted cow skeletons.

Main Results:

• The YOLOv4 model achieved 94.58% average precision (AP) for multi-cow detection.
• The algorithm demonstrated high AP for leg keypoints (91.6%) and varying AP for other body parts under different conditions (day/night, single/dual cows, occlusion).
• Pose classification achieved high precision: 91.67% for standing, 92.97% for walking, and 99.23% for lying.

Conclusions:

• The proposed multicow pose estimation algorithm offers a reliable method for monitoring dairy cow status and well-being.
• The system provides a valuable theoretical reference for large-scale precision livestock farming applications.
• The algorithm's robustness across different conditions suggests its practical utility in real-world farm settings.