Privacy-Preserving Non-Contact Sleep Monitoring via Multimodal Thermal-Depth Sensing

Sleep posture and movement reflect sleep quality and health. Polysomnography (PSG), the clinical standard, is costly and impractical for long-term home use due to its facility needs and contact-based sensors. Thermal infrared array (IRA) sensors offer a low-cost, privacy-preserving alternative, but their low resolution, blanket occlusion, and heat residual limit accuracy. SomnoSense, our non-contact, privacy-preserving sleep monitoring system, integrates IRA thermal data with low-cost direct time-of-flight (dToF) sensor depth data to form a 3D body profile. A lightweight fusion module, combining a random forest classifier and a state machine, addresses blanket occlusion and heat residual errors. Deep neural networks then enable posture recognition and motionlevel estimation. Evaluated on a dataset of 170 minutes of synchronized IRA and depth recordings from five subjects across six postures and three environments, SomnoSense achieves 98.7% detection accuracy, 96.1% posture recognition accuracy, and a mean absolute error value of 0.718 for standardized motion levels. A one-week, 45-hour case study further demonstrates its long-term monitoring capability, revealing distinct sleep posture and motion patterns. These results highlight SomnoSense's potential for practical sleep monitoring in real-world home environments.