Signal quality detection towards practical non-touch vital sign monitoring
Non-touch vital sign sensing is gaining popularity because it does not require users' cooperative efforts (e.g., charging, wearing) thus convenient for longitudinal monitoring. In recent radio-based heart and respiration rate (HR and RR) sensing using Wi-Fi, millimeter wave (mmWave), or ultra-wideband (UWB), inevitable user movements or background moving objects cause large disturbances to the much weaker respiratory and heart signals. Such "corrupted"signals must be detected and excluded to avoid making erroneous measurements. Despite several attempts, reliable signal quality detection (SQD) remains unresolved. In this paper, we spent over 80 hours to manually examine 50268 data samples collected from 8 participants. We find that heart and respiration signals are not always simultaneously available, which breaks an important assumption in prior work. We propose a 2-bit SQD to classify their "availability"separately. We further quantify the contributions of and correlation among a comprehensive set of features in both time and frequency domains, and use a forward selection strategy to identify an optimal and much smaller feature set for multiple common classification algorithms. Extensive experiments show that our 2-bit SQD achieves 91/95% precision, 88/91% recall in detecting available RR/HR signals, as compared to a flat spectrum detector (FSD)  and a spectrum-averaged harmonic path detector (SHAPA)  in prior work, and reduces the 80-percentile RR/HR errors from 10/18 bpm to 3.5/4.0 bpm, 3∼4 fold reductions.