Radar Heartbeat Monitoring with Transformers

Hybrid radar signal processing + Transformer fusion for robust heartbeat/HRV under random body motion (RBM)

Overview

This is an ongoing, funded project developed for IEEE AESS Challenge Problem I: Radar-Based Heartbeat Monitoring in Dynamic Scenarios. The goal is robust estimation of average heart rate (AHR) and heart rate variability (HRV) from FMCW radar, even when the subject exhibits random body motion (RBM), which can easily overwhelm the much smaller cardiac micro-motions.

Our approach is a hybrid pipeline: conventional radar processing to stabilize and isolate the chest signal, followed by a multi-stream Transformer that fuses physics-guided features and produces (i) a super-resolved ECG-like waveform and (ii) AHR/HRV metrics aligned with the official challenge scoring.

Technical Approach (Proposal Summary)

1) Radar processing block (stabilize the chest signal)

Starting from raw FMCW radar data, the signal-processing front-end forms range/slow-time representations, removes clutter, detects the chest range cell, and compensates for range migration so that downstream learning sees a stabilized time series.

2) Physics-guided feature construction (multi-stream time series)

From the stabilized chest signal, we construct a feature tensor that includes complementary views of motion:

  • Displacement slow-time series R[k]
  • Increment ΔR[k]
  • Amplitude at the chest bin and neighboring range bins
  • Band-pass filtered views of R[k] for respiration [0.2, 0.5] Hz, heartbeat [0.7, 2.0] Hz, and heart sound [20, 200] Hz
  • A coarse RBM indicator derived from spectral energy spreading

This is designed so each stream is interpretable on its own, while still giving the model enough context to handle motion corruption.

3) Multi-stream Transformer fusion + temporal super-resolution

The model concatenates the streams into a multidimensional time series x[k], adds positional encoding, then uses a multi-head Transformer decoder with cross-attention to dynamically re-weight streams under RBM (e.g., down-weight corrupted displacement while attending to the RBM indicator).

4) Outputs aligned with the challenge metrics

The network outputs: 1) a reconstructed ECG-like waveform highlighting heartbeat micro-motions, and
2) cardiac metrics: AHR and HRV, matching the official challenge evaluation.

Robustness via Synthetic RBM Augmentation

RBM is treated as the dominant interference source, so the proposal includes a targeted augmentation strategy: generate synthetic RBM displacement sequences using a Markov stochastic process, then add them to the clean displacement signal to increase training diversity.

Two parametric RBM models are described:

  • Random-walk RBM (piecewise-constant velocity subframes + noise) for abrupt shifts/jerks
  • AR(1) RBM for band-limited, temporally correlated jitter

These augmentation parameters can be treated as tunable training hyperparameters.

My Implementation Work (In Progress)

My contributions are focused on the training/experimentation infrastructure and the end-to-end ML pipeline:

Data pipeline (raw radar → model-ready tensors)

  • Implemented the data flow from raw radar captures through a clean Dataset abstraction and DataLoader abstraction
  • Modular feature construction so the same dataset can feed multiple model families (baselines + Transformer variants)
  • Reproducible preprocessing with explicit configuration of CPI/windowing, stream definitions, and augmentation toggles (real + synthetic RBM)

Training + experimentation framework

  • Prototyped a general ModelTrainer that standardizes:
    • logging/metrics per CPI
    • checkpointing
    • evaluation hooks consistent with the challenge deliverables
  • Integrated with Ray Tune to run hyperparameter sweeps (model size, attention settings, augmentation parameters, optimization schedule)
  • Support for multi-device training and controlled experimentation (repeatable runs, comparable configs across models)

Challenge Context & Organizers

This work targets IEEE AESS Challenge Problem I: Radar-Based Heartbeat Monitoring in Dynamic Scenarios. Official pages:

The challenge problem itself was proposed/developed by Yu Rong, Daniel Bliss (Arizona State University), Kawon Han, and Christos Masouros (University College London). ieee-aess.org

Status

  • Funded through the IEEE AESS Challenge program; currently under active development and evaluation against the official scoring pipeline.