Highly Generalized Sleep Posture Recognition Using FMCW Radar

Identifying users’ sleep posture is significant in reducing sleep apnea events and avoiding postoperative pressure sores. Past studies have identified sleep posture by installing cameras, installing sensors on mattresses, or letting users wear wearable devices. However, the camera-based method is usually affected by light intensity or coverage and can invade users’ privacy. The method based on contact sensors will affect the comfort of users’ sleep. The use of radar can solve the problem of cameras and contact sensors, but previous studies need to collect data from new users for calibration to maintain high performance. In this work, we use FMCW radar to estimate the position image of the user in space. We propose a multi-task learning sleep posture recognition model based on a neural network, which uses the radar position image to estimate the user’s sleep posture. In addition, we use mix-up data enhancement to improve the model’s generalization. We collected data from 17 subjects to train and test our model. The proposed method can achieve 0.935 sleep posture recognition F1 score without collecting new user calibration data.