An AI-based real-time framework for multi-behavior detection and tracking of Mithun (Bos frontalis) using YOLOv8n and deepSORT

Abstract

Mithun (Bos frontalis), a culturally and economically significant semi-domesticated bovine species native to the hilly regions of Northeast India, plays a vital role in tribal livelihoods and regional ecosystems. Traditional methods for monitoring Mithun behavior rely on manual observation, which is labor-intensive, time-consuming, and unsuitable for continuous surveillance. To address this limitation, we propose the first artificial intelligence-based, real-time, non-contact framework for automatic behavior detection and tracking of Mithun in natural farm environments. The system integrates the YOLOv8n object detection model with the DeepSORT tracking algorithm to recognize four key behaviors: feeding, standing, lying, and mounting. A custom dataset of 3,000 manually annotated images was developed from high-resolution CCTV footage captured under varied lighting and environmental conditions, including infrared night vision. The trained model demonstrated strong performance with a mean average precision (mAP@0.5) of 0.995, a recall of 0.996, and a real-time inference speed of 31 frames per second. Extensive quantitative evaluations, confusion matrix analysis, and qualitative visualizations confirmed the model’s robustness under occlusion, low light, and background clutter. DeepSORT was incorporated to support qualitative identity-aware tracking across frames; however, tracking performance was not quantitatively evaluated. This scalable, non-invasive framework offers significant potential for improving reproductive management, enhancing animal welfare, and enabling continuous precision livestock monitoring, particularly in remote or resource-constrained tribal farming systems.