Instance segmentation and automated pig posture recognition for smart health management

Area of Science:

• Agricultural technology
• Computer vision
• Animal welfare science

Background:

• Monitoring pig posture and movement is crucial for detecting health issues and abnormal development in growing pigs.
• Manual monitoring in large-scale farms is labor-intensive and time-consuming.
• Computer vision offers a potential solution for automated pig monitoring.

Purpose of the Study:

• To develop and evaluate a computer vision system for automated pig posture recognition and detection.
• To utilize masked-based instance segmentation for monitoring pigs in a closed farm environment.

Main Methods:

• Two video acquisition systems (top and side views) were used to capture RGB images.
• A dataset of 600 manually annotated images was created, classifying four postures: standing, sitting, lying, and eating.
• The Mask R-CNN framework was employed for instance segmentation, including region proposal network, RoIPool, classification, and bounding-box regression.

Main Results:

• The model achieved high accuracy in identifying pig postures, with a mean average precision of 0.937 for piglets and 0.935 for adults.
• The system demonstrated potential for real-time monitoring and early detection of welfare issues.
• Body weight estimation using 2D image pixel area showed a high correlation with actual weight.

Conclusions:

• The proposed computer vision model is effective for automated pig posture monitoring, supporting early welfare issue detection and farm management optimization.
• Further research incorporating 3D imaging and testing in diverse farm conditions is recommended to enhance real-world applicability.