When to Cancel the Walker First: Successive Motion Cancellation for Radio Vital Signs

Radar and Wi-Fi can measure breathing and heartbeat without touching the body, but only while the person stays still. Breathing moves the chest by a few millimetres and a heartbeat by much less, while walking moves the whole body by tens of centimetres. The reflection is therefore dominated by the motion, and because walking cadence sits inside the heart-rate band, frequency filters cannot pull the two apart. Wireless communications faces a structurally similar problem in non-orthogonal multiple access (NOMA). There, a base station sends a loud signal and a quiet signal on the same frequency at the same time. The receiver decodes the loud one, subtracts it, and reads the quiet one from what remains, a procedure called successive interference cancellation (SIC). NOMA theory says this works only when the two signals differ strongly in power. It also says exactly how errors in the first step leak into the second.

We propose bringing that second piece of NOMA theory, the error-propagation analysis and the rule for when to pair, into radio vital-sign sensing. The system first estimates the body's gross motion with an explicit, low-dimensional model. It then removes that motion from the radar phase and estimates breathing or heart rate from the residual. A disparity criterion, computed from the first stage alone, decides whether this ordered cancellation should be trusted or whether the system should fall back to joint separation. The claim to test is that the criterion predicts where cancellation succeeds and where it fails, and that switching on it beats both always-cancel and the current one-shot learned disentanglers.

Core Claim

For a single dominant reflector, the radar phase is approximately $\phi(t) = \frac{4\pi}{\lambda}\big(d_m(t) + d_v(t)\big)$. Here $\lambda$ is the wavelength, $d_m$ is gross-motion displacement and $d_v$ is cardiopulmonary displacement. The two add linearly in unwrapped phase. The nonlinearity that MoRe-Fi emphasizes comes from phase wrapping, range-bin migration, and multipath.

Suppose a first stage estimates $\hat d_m$ and subtracts it. The residual is $d_v + e$, where $e = d_m - \hat d_m$. Write $\rho = \sigma_m/\sigma_v$ for the disparity ratio between motion and vital amplitudes, and $\varepsilon = \sigma_e/\sigma_m$ for the first stage's relative error. The vital signal then sees an effective residual signal-to-interference ratio

$$ \mathrm{SIR}_v = \frac{1}{(\varepsilon\,\rho)^2}. $$

This is the direct analogue of imperfect-SIC analyses in NOMA scheduling, where residual interference is a fraction of the strong user's power (Celik et al., Celik et al.).

The expression predicts a U-shape rather than "more disparity is better":

  • Small $\rho$ (fidgeting, typing). The motion estimator cannot distinguish motion from breathing, so $\varepsilon$ is large.
  • Very large $\rho$ (fast walking). Even a small, model-mismatch-limited $\varepsilon$ leaves a residual that swamps a millimetre-scale breath, let alone a heartbeat.

This mirrors NOMA's finding that pairing users of similar strength gives diminishing returns (Zeng et al., de Sena et al.), while sufficient disparity yields gains for both parties (Ding et al.). The new contribution is not the subtraction itself. It is a predictive, measurable quantity, $\varepsilon\rho$ estimated online from the first stage, that tells the system when ordered cancellation is the right decoder.

Relation to Prior Work

Two lines of work motivate this proposal and supply the competitors.

On the sensing side, two systems remove motion inside a single learned model:

  • RF-vital trains a U-Net encoder with an adversarial head that penalizes any recoverable walking trajectory.
  • MoRe-Fi aligns I and Q latent distributions with a Wasserstein penalty in a variational encoder-decoder.

Neither makes the motion estimate explicit, orders the decoding, or models how motion-estimation error propagates. We do all three. We treat these systems as the strongest baselines rather than as straw men.

Wital takes a different route. It uses Ricean K-factor analysis to show that a strong direct path compresses the dynamic range available to chest reflections. It handles macro-motion by detecting and discarding irregular windows, which constrains our design: discarding is a legitimate baseline, but it cannot provide continuous daytime monitoring.

On the communications side, NOMA-empowered integrated sensing and communication is the closest bridging work, and both papers model the radar target as a point:

  • NOMA Empowered ISAC multiplexes users in the power domain when antennas are scarce or channels are correlated, reporting a 2–3 dBm effective-sensing-power gain in overloaded settings. Its targets are point scatterers with perfect channel knowledge.
  • NOMA Inspired Interference Cancellation for ISAC treats the radar waveform as a "virtual user" that communication receivers cancel by SIC.

Both perform SIC at communication receivers on point-target waveforms. We perform it inside the sensing receiver, on a body whose echo carries both macro-motion and micro-Doppler physiology. We borrow two specific ingredients from NOMA theory: the error-propagation model, and the pairing logic from Ding, Adachi and Poor. We do not borrow power allocation, because the sensing receiver does not control how loudly the body moves.

Research Program

Stage 1: explicit motion estimate. Stage 1 tracks the strongest body reflector across range bins with a 60 GHz FMCW radar, the same class of hardware as Soli. It fits a smooth parametric trajectory: a low-order spline in range plus a periodic gait term fit by short-window harmonic regression. The fitted trajectory is reconstructed as a complex phasor, and the system multiplies the raw I/Q stream by its conjugate. Doing this in the complex domain sidesteps wrapping errors. Stage 1 also outputs its own uncertainty, estimated from held-out fit residuals and bootstrap over windows. That uncertainty gives an online estimate of $\varepsilon$.

Stage 2: vital-sign estimation from the residual. Stage 2 is a small residual-aware network that estimates breathing and heart rate. It is trained on residuals produced by Stage 1, not on clean data, so it learns the specific statistics of imperfect cancellation. This is the sensing counterpart of designing a NOMA decoder for imperfect SIC.

The switch. A rule compares the online $\hat\varepsilon\hat\rho$ to a threshold calibrated on training data. Below the threshold, the system uses the cancel-then-estimate output. Above it, the system uses a joint estimator, a MoRe-Fi-style disentangler. This is the analogue of NOMA falling back to orthogonal access when pairing does not pay.

Multi-subject extension. The same logic carries over to two people in the overloaded, correlated-channel regime identified by NOMA Empowered ISAC. Consider a walking person and a seated person at similar angles, with fewer antennas than reflectors. The walker's reflection is the strong signal and is cancelled first, which exposes the seated person's breathing. Ordering follows reflected-power disparity, as in NOMA user pairing. This addresses the multi-person separation that Wital leaves open, using a different mechanism from the correlation-and-phase clustering of MultiResp.

Baselines

Naive reference. A Butterworth band-pass filter on the extracted phase, followed by spectral peak picking. Both Soli and MMPD show this kind of approach failing under motion; on MMPD, POS rises from 9.70 to 17.05 bpm MAE when subjects walk. It sets the floor.

Gated discard. Following Wital, this baseline applies regularity detection and reports vitals only in windows judged stationary. It will be scored on accuracy and on coverage, the fraction of time with a valid estimate. Beating it means matching its in-window accuracy while reporting during motion.

Strongest competitors. These are RF-vital's multi-task adversarial network and MoRe-Fi's IQ variational encoder-decoder. Both will be reimplemented and trained on exactly the same recordings, splits, and input representation (complex I/Q range profiles). They get the same GPU budget as our Stage 2 and an equal hyperparameter-search budget of trials.

Competitor Reported result Dataset Directly comparable to our recordings?
RF-vital 3.67 BPM respiration MAE during active walking RRM-moving No
MoRe-Fi Average cosine similarity 0.9162; instantaneous respiration-rate error mostly below 0.1 bpm MoRe-Fi custom UWB dataset No
Soli 1.69 bpm heart-rate MAE Sleep Dataset Low-motion reference only

None of these numbers is on our recordings, so the comparison that counts is the controlled reimplementation. If the original RRM-moving data can be obtained, we will also report there against the 3.67 BPM figure. Beating these competitors means:

  • lower respiration MAE overall, with a paired bootstrap confidence interval excluding zero; and
  • a larger margin on held-out motion types (motion types never seen in training) than on in-distribution motion.

The second condition is the regime where an explicit physical motion model should help most.

Ablations isolating the key ingredients. Four ablations each target one component:

  • Stage 1 followed by the band-pass filter, with no learned Stage 2. This is classical motion compensation, and it isolates the value of a residual-aware second stage.
  • Always-cancel and always-joint variants, which isolate the value of the switch.
  • An oracle switch, which bounds how much the switch could possibly gain.
  • An end-to-end network given the Stage 1 trajectory as an extra input channel. This separates "ordering the decoding" from "having extra information."

Why not just…?

Why not just train a bigger end-to-end disentangler on more motion data? This is the real threat. RF-vital and MoRe-Fi already learn motion removal implicitly, and with enough data they may learn the U-shape on their own. We do not argue this away. The extra-input ablation and the held-out-motion split settle it. If the end-to-end model, given the same Stage 1 information, matches the switched system on unseen motion types, the contribution reduces to the diagnostic criterion, and we will report it that way. Our prediction is that it will not match: an explicit trajectory model degrades gracefully on new gaits, while a learned encoder trained to purge seen gaits has no such guarantee.

Why not just use classical motion compensation (track the body, subtract its phase, then filter)? This is our Stage 1 followed by a band-pass filter, and it is run as an ablation. The imperfect-cancellation analysis predicts it fails at both ends of the disparity range. At small $\rho$ it subtracts breathing along with motion. At large $\rho$ it leaves a residual larger than the heartbeat. The disparity sweep below will show whether that failure is real and whether the residual-aware Stage 2 and the switch repair it.

Why not just discard motion windows? Wital shows discarding works for rolling over in bed. During walking or typing, however, it yields no estimate at all. Clinical work on ambulatory monitoring (Ambulatory AF Monitoring with Wearable PPG) stresses that such gaps miss events. The coverage metric makes this cost visible.

Why not just separate people with more antennas? NOMA Empowered ISAC shows that spatial separation breaks down when reflectors outnumber antennas or sit at correlated angles. Our two-person experiment deliberately places subjects in that regime and includes a beamforming-only baseline using the same array.

Proposed Experiments

Disparity sweep on a phantom. This is the decisive check. A motorized linear stage carries a voice-coil actuator. The stage produces macro-motion of controlled amplitude and spectrum, including gait-like periodic motion overlapping 1–2.5 Hz. The actuator produces breathing-scale and heartbeat-scale displacement. Because ground truth is exact, $\rho$ can be swept over several orders of magnitude. We will measure Stage 1's $\varepsilon$, plot vital-rate error against $\varepsilon\rho$, and test whether the curves for different motion types collapse onto one predicted relation. If they do not collapse, the core claim fails.

Human recordings. We will record 60 GHz FMCW data from human participants with chest-belt and ECG ground truth. The protocol covers seated rest, typing, fidgeting, treadmill walking at several speeds, and turning. Primary metrics are respiration MAE, heart-rate MAE, and waveform cosine similarity, chosen to match MoRe-Fi's reporting, plus coverage. Heart rate under walking is treated as a stretch goal, since the analysis predicts it lies past the right arm of the U.

Two-person scenes. A walker and a seated subject share an angular sector, with an antenna count below the number of significant reflectors. We report the seated subject's respiration MAE with walker-first cancellation, with beamforming only, and with the reimplemented joint disentangler.

Success criteria. The program succeeds if all three of the following hold:

  • the phase-domain error model predicts the crossover point between cancellation and joint estimation to within a factor of two in $\rho$;
  • the switched system beats both reimplemented learned baselines on held-out motion types; and
  • walker-first cancellation recovers the seated person's breathing where beamforming alone does not.