FMCW Radar-based Sleep Posture Recognition with Transfer Learning and Range-Aware Dataset

Sleep posture recognition is crucial for understanding sleep quality and diagnosing sleep-related disorders. In this paper, we propose a fine-tuned VGG16 model for classifying four fundamental sleep postures—Supine, Prone, Right side, and Left side—using data collected from Frequency Modulated Continuous Wave (FMCW) radar. Radar signals are processed into range-time images and fed into the deep-learning model for training and classification. We evaluate the impact of varying window sizes, radar distances (1.5, 1.7, and 1.9 meters), and subject independence on the model’s performance. Our fine-tuned VGG16 model achieves an average accuracy of over 96% across different distances and demonstrates superior performance compared to a basic CNN. However, the model struggles to generalize to unseen subjects, indicating the need for further research to address subject-dependent challenges. This study highlights the potential of radar-based systems for contactless sleep monitoring and suggests future improvements to enhance model generalization.