mmWave-Diffusion:A Novel Framework for Respiration Sensing Using Observation-Anchored Conditional Diffusion Model
A new AI framework transforms noisy radar signals into accurate breathing data in just 20 steps.
A research team led by Yong Wang has introduced mmWave-Diffusion, a novel AI framework that significantly improves the accuracy of contactless respiration monitoring using millimeter-wave (mmWave) radar. The core innovation is an "observation-anchored conditional diffusion model" that directly models the residual between noisy radar phase observations and the true respiratory signal. Unlike standard diffusion models that start from pure noise, this approach initializes the generative sampling process within a neighborhood consistent with the actual radar measurements. This alignment with measurement physics drastically reduces the computational overhead, requiring only 20 reverse diffusion steps for robust inference.
The framework is powered by a custom architecture called the Radar Diffusion Transformer (RDT). The RDT is explicitly conditioned on the raw phase observations and enforces strict temporal alignment through patch-level dual positional encodings. Crucially, it injects local physical priors using a banded-mask multi-head cross-attention mechanism, which helps the model distinguish subtle breathing patterns from non-stationary interference like body micromovements. Evaluated on a substantial dataset of 13.25 hours of synchronized radar and respiration data, the system achieves state-of-the-art performance in both waveform reconstruction and respiratory-rate estimation, demonstrating strong generalization capabilities. The work has been accepted for presentation at the prestigious IEEE ICASSP 2026 conference.
- Uses a novel observation-anchored conditional diffusion model to remove body motion interference from radar signals, starting inference from measurement-consistent data instead of random noise.
- The custom Radar Diffusion Transformer (RDT) enables robust denoising in just 20 reverse steps by using dual positional encodings and banded-mask cross-attention to inject physical priors.
- Achieved state-of-the-art respiration monitoring results on 13.25 hours of real data, enabling fine-grained, contactless health sensing with strong generalization.
Why It Matters
Enables accurate, non-invasive continuous health monitoring for sleep studies, patient care, and wellness tracking without wearable sensors.