Research outcomes

---

Experimental environment

---

...

Development of analog computational architecture

---

FMCW radar delivers data in sweeps of some fixed length, which contain information on spatial, instantaneous structure of monitored environment. Distance to objects is encoded in signal frequency components, so to reconstruct this structure, one needs to analyze frequency spectrum of each sweep. Detection of vital signs is feasible through identification of characteristic displacements of object surfaces – chest motion that reflects breathing or surface vessel inflation caused by blood pressure waves. Vital signs detection and estimation of their parameters can thus be accomplished by applying signal processing procedures that attempt to identify these micro-motion in signal frequency domain. A standard way to solve this problem is to apply double-DFT procedure. Its first phase is DFT performed on a single sweep signal and its objective is to detect spectral local maxima, which indicate presence of objects in corresponding range-bins from a radar (for example, for the considered radar, presence of a local maximum in magnitude sweep spectrum at 1 kHz frequency indicates an object located at approximately 1.4 meters from a radar). The second phase, called slow-time DFT is a spectral analysis performed only on a sequence of spectral components taken from subsequent sweeps at the position of interest (e.g. 1kHz), and its objective is to identify phase-shifts that correspond to micro-motions of an object at the considered distance. Although the presented procedure works well in detecting e.g. breathing rate, it does not fit into analog computing framework, as all analyses involved are digital. Therefore, an objective of our research was to develop fully analog procedure that provides an alternative way of the task realization.

Block diagram of the elaborated computational architecture has been presented in the figure below. The presented analog signal processing algorithm operates on both In-phase and Quadrature components, both produced by FMCW radar. A key operation that enables extraction of micro-motion related component from radar information is mixing of both signals from a current sweep with a delayed (by one sweep) version of In-phase component. Low-pass filtering of the resulting waveforms leaves information only on between-sweep phase-shifts (wrapped by sine and cosine functions), which can be subsequently reconstructed and used as a domain for parameter estimation.

Results of Breathing Rate estimation for different experimental scenarios are shown in Fig. 2, where estimation results produced by a reference digital signal processing algorithm (blue color) and the proposed analog method (plot in orange) are almost indistinguishable. A reference information, produced by a ZephyrTM tensometric belt (shown in green) are delayed with respect to instantaneous breathing rate by approximately 30s integration time are also matching closely the presented estimates (with exception at the beginnings of new experimental scenarios, marked using vertical red lines) where reference information is useless due to lack of sufficient amount of data).

Machine learning algorithms for data analysis

---

Machine learning provides means to cope with unpredictable and extremely complex real-world contexts for vital sign detection and estimation of vital sign parameters. Radar-based monitoring of closed spaces results in highly complex, noisy signals with multiple reflections, low spatial resolution and poor directional sensitivity, making analytical handling of data almost impossible. At the same time, such cluttered environments with several potential sources of mechanical noise (such as e.g. operating fans) are one of primary application targets. Therefore, one of the main objectives of our research was to develop machine learning methods for person presence detection, as well as for estimation of basic parameters of liveness-indicators, such as e.g. a breathing rate.