Recognition and monitoring of human motions using RF signals

In this thesis, the potentials of radio frequency (RF) signals are investigated for indoor and through-the-wall (TTW) health monitoring and activity recognition applications. The moving or stationary objects on or near the propagation path of electromagnetic waves affects the received RF signals waveform. Recently, these RF signals are examined in literature for different healthcare applications, such as fall detection, vital sign monitoring, etc., and activity recognition. It can be life-saving to monitor the respiratory rate (RR), one of the vital signs, even for healthy people in real-time. It is reported that coronavirus disease 2019 (COVID-19) causes mild respiratory symptoms in the early stage. It is more important to continuously monitor the RR of people in nursing homes and houses with noncontact methods. In this thesis, the noncontact vital RR monitoring is chosen as the first application and three different signal models and new methods are proposed. In the proposed system, the high resolution subspace-based parametric spectral estimation approaches, estimation of signal parameters by rotational invariance technique (ESPRIT) and multiple signal classification (MUSIC), and auto-regressive and moving average (ARMA) model-based techniques are presented as the RR estimation algorithms. It is shown by the experiments with real-world settings and measurements that the proposed noncontact RR monitoring system achieves very accurate results with the limited number of observations and outperforms the periodogram method commonly used as the benchmark in the literature. The standard joint unscented Kalman filter (JUKF) method is also modified for this new and time-critical problem and it is shown in the experiments that the proposed modified JUKF (ModJUKF) method attains a low error rate according to the windowing-based methods in the time-varying RR scenario. As the second application, TTW static/moving human detection problem is investigated. The presence detection of the static human behind the wall using minute reflections caused by human breathing is the first time proposed/used in the literature. It is shown in various experiments that the proposed machine learning-based method achieves over 99% accuracy for the detection of both static and moving human in double TTW scenario. It is also validated with the real experiments that the proposed system can be used in a new environment without retraining the machine learning model. The last application is a human-machine interaction application which is a device-free air-writing recognition system that classifies 26 capital letters using RF signals. The proposed system uses the Discrete Cosine Transform (DCT) coefficients as discriminative features for the first time for gesture recognition. The oppositely polarized (i.e. horizontal and vertical) antennas are used in the two-channel receiver to provide polarization diversity that is exploited to improve the classification accuracy. It is shown with experiments conducted with real measurements that the proposed system, which achieves 95.15% accuracy in the classification of 26 air-written letters, outperforms the fairly new WiFi-based air-writing recognition approaches. In all these applications, RF signals are generated and captured by low-cost software-defined radio (SDR) modules. The presented systems are validated through various experiments conducted with real measurements collected from different numbers of volunteers and from different places compatible with realistic scenarios.