Description of Invention
This project presents a sleep posture recognition system using a pretrained Vision Transformer (ViT) with transfer learning. Traditional methods are costly and struggle with occlusions like blankets. Our approach utilizes the SLP dataset, covering three posture categories under varying blanket conditions. ViT’s global attention mechanism enables accurate posture classification, even with thick blankets. The best-performing model, ViT-B/16@384, achieved 99.36% testing accuracy. Therefore, this work demonstrates strong potential for sleep monitoring and healthcare applications. Future improvements include expanding the dataset and enabling real-time monitoring for practical deployment in clinical and home environments.