MULTI-OBJECT TRACKING AND VISION TRANSFORMER ENHANCEMENTS FOR REAL-TIME COW MONITORING: A REVIEW AND IMPLEMENTATION STUDY

Abstract

n recent years, precision livestock farming has emerged as a vital area of innovation, aiming to enhance the health, welfare, and productivity of animals through advanced technologies. This thesis investigates deep learning-based techniques for cow detection, tracking, and behavioral monitoring by integrating insights from a comprehensive literature review and a practical implementation. The review explores the evolution of object detection and tracking models, such as YOLO, R-CNN, and Vision Transformers, evaluating their applicability under varying farm conditions including occlusion, lighting challenges, and dense cattle populations. Building upon this foundation, the thesis proposes a novel hybrid system—YOLOv8s augmented with Coordinate Attention (CA) and integrated with DeepSORT and a Vision Transformer (ViT)—designed for accurate cow face detection and re-identification. The model was trained and tested on a custom dataset collected from a farm in Uttar Pradesh, India, comprising over 400 hours of video. Evaluation metrics show that the proposed system achieved an IDF1 score of 92.5%, MOTA of 88.4%, and MOTP of 97.2%, outperforming traditional methods such as ByteTrack, BoT-SORT, and DeepSORT. Additionally, the system demonstrated a 50% reduction in identity switching and a 20% improvement in processing time. These findings underscore the potential of the proposed approach to serve as a reliable solution for real- time livestock monitoring, contributing to smarter and more efficient dairy farm management.