A Two-Stage Computer Vision Framework for Individual Recognition and Precision Data Collection Illustrated through Piglets in Farrowing Pens

Intensive livestock farming practices in the modern era are characterized by large flock sizes and high animal densities. These conditions make traditional manual observation methods inefficient, impractical, and error-prone. While computer vision-based research has made notable progress in studying animal populations, it often falls short in enabling long-term, precise individual recognition for comprehensive behavioral analysis and individual health management. This limitation hampers in-depth research on animal physiology, nutrition, health, behavior, and welfare. In this study, we propose a two-stage individual recognition and tracking scheme for piglets in commercial farrowing pen settings. Our approach provides a valuable tool for pig research by continuously recognizing and tracking individual piglets, accurately measuring their behavioral responses to different experimental treatments. Furthermore, this method can be adapted to other species as well. In the first stage, we employ a lightweight model based on improved YOLO to detect the presence of the sow and piglets within the farrowing pen. In the second stage, we utilize a fine-tuned handwritten digit recognition model based on color-adjusted Mnist to identify the colored digits (e.g., digits 0 to 3 in blue, green, and red) on the bodies of piglets. By combining the two stages, our approach enables individual piglet detection and identification. Experimental results demonstrate an average recognition accuracy of 97.6%, successfully identifying 12 piglets over long periods. In practical applications, this approach has the potential to greatly facilitate animal science research. © 2024 11th European Conference on Precision Livestock Farming. All rights reserved.