The Problem

Continuous vital sign monitoring (heart rate and breathing rate) is a clinical necessity in hospitals, ICUs, and care wards because it enables early identification of subtle health changes, improves early recognition of patient deterioration, and helps detect clinical deterioration sooner. The standard tools are contact based: ECG electrodes, pulse oximeters, and chest straps, and while these systems may use wearable sensors or non-contact technologies like radar, wearables can limit comfort and mobility and increase the need for manual checks. They work, but prolonged skin contact causes irritation and pressure injuries, adhesive electrodes can directly cause hospital-acquired infections, and patients frequently dislodge sensors at the moments monitoring matters most.

Contactless monitoring using millimetre-wave radar is a better alternative and solves this directly. A radar operating in the 57–64 GHz band uses electromagnetic waves to detect chest-wall displacement as small as 0.3 mm, enough to capture both breathing and heartbeat signatures and monitor human vital signs with no physical contact, no adhesives, and no infection risk.

Across related fields, other non-contact systems leverage radio frequency radar, Wi-Fi, or video analytics to detect movements. Video-based monitoring extracts vital signs using remote photoplethysmography, while audio signal processing can measure respiratory sounds and heartbeats.

Millimetre-wave radar sensors capable of vital sign monitoring are available from multiple vendors including Infineon (BGT60TR13C), Texas Instruments (IWRL6432), and Acconeer (XM125), and their reference SDKs and demos typically demonstrate vital sign monitoring for one person only, despite broader applications in medicine and smart homes, where some platforms can monitor heart rate, respiratory rate, and blood pressure. Continuous remote monitoring also helps reduce cross-infection risk and the time nurses spend on manual vital sign checks. Real clinical environments like ICU bays, general wards, post-operative rooms routinely have two or more patients in view. This project builds on existing single-person mmWave reference implementations and extends the approach to simultaneous dual-person monitoring, with all the processing running on the chip itself.

What was Built? (Using FMCW Radar)

The work was split into two parts: building a vital signs extraction module with the core system components of data capture and signal processing and integrating it into the reference SDK’s real-time processing chain running on the chip’s Cortex-M4F processor.

The radar works by transmitting radio waves and recording the reflections that return. The sensor used in this project has two transmitting antennas and three receiving antennas. Every signal transmitted is picked up by all three receiving antennas, giving six unique signal paths in total, so the system can analyze data from multiple sources simultaneously. More signal paths provide better spatial resolution, which is what allows the system to separate two people standing at different positions.

The radar receives signals that contain two pieces of information: signal strength and phase. The phase is what matters here, when a person’s chest moves even by a fraction of a millimetre, it changes the distance the radar signal travels, and that change shows up in the phase. The signal processing pipeline uses this phase change to calculate exactly how much each person’s chest and body have moved. That chest displacement data is then processed to extract breathing rate and heart rate, including respiration. Conventional filter-based methods separate vital signals using frequency components, but more advanced methods such as DWT and VMD can improve separation in dynamic conditions. EMD also has limitations because mode mixing can occur, and selecting the right wavelet basis remains a practical challenge that these methods must address.

Integration into the Reference Software Development Kit (SDK) Processing Chain

The vital signs module does not run on its own. It only activates once the SDK’s built-in people tracker confirms two people in the scene and identifies their positions. Each person’s position is then converted into a radar range, the specific point in the signal where their chest movement appears and passed into the pipeline. This required modifying a core file inside the reference SDK, one that manages the main processing loop of the radar, while ensuring the entire process still completes within the radar’s real-time timing window.

Five Fixes that Made Vital Signs Monitoring Reliable

Getting an initial estimate of breathing rate and heart rate was straightforward, but making the readings reliable under real conditions required additional signal-processing methods.

Five specific problems had to be solved:

The first was stability. If the tracker briefly lost a person and snapped back to the wrong position, the entire signal history would be corrupted. Bad frames were simply discarded, and a confidence threshold was introduced to keep short tracking slips from contaminating the reconstruction.

The second was noise. The system read the signal data from across the full radar range, mixing real chest movement with everything else moving in the room. As with conventional frequency-based filtering used to address poor separation, narrowing the extraction to only where breathing and heartbeat appear fixed this.

The third was false peaks. Without smoothing the signal before frequency analysis, breathing was leaking energy into the heart-rate range and generating peaks that looked like real heartbeats. Smoothing and reconstruction can be combined before spectral analysis to reduce breathing leakage into the heartbeat band. The reconstructed signal was then used to estimate heart rate more cleanly.

The fourth was signal drift. Every time a person shifted position, the internal reference needed a reset. Without it, the displacement measurement carried a false offset that poisoned every reading that followed. The proposed reset logic also kept the downstream spectrum from inheriting stale motion bias.

The fifth was false reading outputs. The system was producing readings even when nobody was present, picking up wall reflections. A presence check tied to the people tracker stopped any output unless a real person was confirmed in the scene, helping confirm that the output corresponded to a real person rather than wall reflections.

Wavelet-based methods reported in prior work use stronger separation and have reached up to 93% accuracy for heartbeat monitoring, which helps explain why this kind of processing matters here.

Observations on Heart Rate and Breathing Rate

The system was capable of consistently monitoring vital signs from two people in a controlled indoor environment, and the results demonstrate stable monitoring in this state. Both subjects produced stable and reliable breathing rate readings. Heart rate extraction was more difficult, as the chest displacement caused by a heartbeat is significantly smaller than that caused by breathing, making respiratory rate motion harder to separate from noise. Two-person separation performed well when subjects were positioned at sufficiently different angles relative to the radar, as shown by the individual signal extraction confirmed for both subjects. Movement or breath-holding by one subject had no measurable effect on the readings of the other. These results support applications where continuous monitoring may improve early recognition of patient deterioration.

Know More@ https://www.einfochips.com/domains/medical-devices/

Frequently Asked Questions

Q.1. What happens if a third person walks into the room during monitoring. Does the whole system break or just ignore them?

Answer: Nothing breaks, the system just ignores the third person. It’s built track only for two persons at a time (two fixed slots: Person 0 and Person 1). If a third person enters while both slots are already get detected by the radar but are never assigned a slot. No vital signs get calculated for them, and no error or crash occurs. Person 0 and Person 1 continue being tracked normally, completely unaffected by the third person’s presence. The system simply has no room for a third identity, it is a capacity limit, not a failure.

Q.2. Was the system tested with subjects in different postures?

Answer: The system was experimentally validated using two upright postures: standing and seated on a chair. These postures were tested separately, and in both cases the radar achieved reliable person detection, target tracking, and vital-sign estimation. Consistent performance was observed because the subjects’ chest region remained within the radar’s configured field of view and operating range, allowing accurate signal acquisition and processing.

Q.3. Can this existing system scale beyond two people?

Answer: Yes, this technology can scale beyond two people, and it depends on a few key parameters. The main one is angular and distance separation. The radar tells people apart based on how far apart they are in distance and angle, so the more people in the scene, the more precise they need to be spread out for the radar to distinguish them individually. Processing power is another factor, since each additional person requires their own signal processing chain running within the same real time window. Available memory also scales with the number of people being tracked simultaneously. In short, scalability is governed by spatial separation between persons, computational capacity, and memory, not by any fundamental restriction of the radar technology itself.