Leveraging YOLOv8 and transfer learning for real-time classification of horned sheep behavior: a computer vision approach to precision livestock

Abstract

Real-time monitoring of livestock behavior is essential for improving animal welfare and optimizing farm productivity. Traditional manual observation is often inefficient, error-prone, and labour-intensive. This study proposes a computer vision-based system for real-time classification of horned sheep behaviours, specifically wake, lay, and sleep, using transfer learning with the YOLOv8 segmentation model. The system is designed for top-view camera setups in one-sheep-per-pen farm structures, enhancing visibility and minimizing occlusion. A dataset of over 12,000 annotated images was collected and augmented to train the model effectively. The proposed system demonstrates a classification accuracy exceeding 90% across all behavior classes, with an average detection latency below one second. Experimental results validate the model’s robustness under varied lighting and environmental conditions. This approach enables efficient and scalable behaviour monitoring without requiring additional sensors, offering a practical solution for intelligent farm management and early detection of health-related anomalies.