A comparison between ceiling-mounted FMCW, IR-UWB and Wi-Fi radar for in-bedroom human activity monitoring and sleep interruption detection

Overview§

This paper presents a systematic, controlled comparison of three RF sensing technologies—frequency-modulated continuous wave (FMCW) radar, impulse radio ultra-wideband (IR-UWB), and Wi-Fi sensing—under identical ceiling-mounted deployment conditions. Using synchronized recordings from 20 participants across six room layouts, the authors evaluate all technologies with the same convolutional neural network (CNN) on two tasks: fine-grained 10-class human activity recognition (HAR) and coarse 4-class sleep monitoring. The study provides the first fair cross-technology benchmark in a healthcare-oriented setting.

Experimental Setup§

• Sensors: Ceiling-mounted FMCW radar, IR-UWB radar, and Wi-Fi transceivers (channel state information).
• Dataset: 20 participants, 6 room layouts, synchronized multi-modal recordings.
• Tasks:
• HAR: 10 activity classes (e.g., lying, sitting, walking, bending, etc.).
• Sleep: 4 coarse states (e.g., in-bed, out-of-bed, restless, still).
• Model: A shared CNN architecture applied to each modality's preprocessed input (range-Doppler maps for FMCW, range profiles for IR-UWB, and CSI amplitude/phase matrices for Wi-Fi).
• Evaluation: Cross-subject (leave-participant-out) and cross-room (leave-room-out) setups, reporting macro F1 scores.

Key Results§

• IR-UWB achieves the highest cross-subject HAR performance: 89.0% macro F1.
• FMCW generalizes best to unseen room layouts: 83.8% macro F1.
• Sleep monitoring: All technologies exceed 92% macro F1 in unseen environments, indicating that coarse sleep monitoring is relatively easy for RF sensing.
• The paper identifies a fundamental trade-off: higher recognition accuracy (IR-UWB) comes with lower environmental robustness, while FMCW offers better generalization.

Explanation of Trade-off§

The authors attribute differences to:

1. Range resolution: FMCW provides higher range resolution, preserving spatial layout information that helps generalization to new rooms.
2. Antenna diversity: IR-UWB's multiple antennas improve spatial diversity for activity discrimination but may overfit to specific antenna placements and room geometry.
3. Doppler resolution: FMCW and IR-UWB offer different sensitivities to body movements, affecting activity granularity.
4. Spatial information retention: Wi-Fi CSI retains coarse-grained environmental signatures, making it robust in known rooms but less discriminative for fine activities.

Methodology Details§

For each modality, raw signals are transformed into time-frequency representations:

• FMCW: Range-Doppler map (RDM), $X_{RDM}[n, m]$, computed via 2D FFT over fast-time and slow-time dimensions.
• IR-UWB: Channel impulse response (CIR) or range profile, processed with a bank of matched filters; a sequence of range bins over time forms a time-range map.
• Wi-Fi: Channel state information (CSI) from OFDM subcarriers; phase and amplitude are cleaned and arranged as a time-series matrix.

The CNN architecture is a small 2D convolutional network (e.g., 4 convolutional blocks with batch normalization, ReLU, and max pooling, followed by global average pooling and a softmax classifier). This is identical across modalities to ensure fairness.

A typical PyTorch implementation for the shared CNN:

import torch.nn as nn

class SharedCNN(nn.Module):
    def __init__(self, num_classes):
        super().__init__()
        self.features = nn.Sequential(
            nn.Conv2d(1, 16, 3, padding=1), nn.BatchNorm2d(16), nn.ReLU(), nn.MaxPool2d(2),
            nn.Conv2d(16, 32, 3, padding=1), nn.BatchNorm2d(32), nn.ReLU(), nn.MaxPool2d(2),
            nn.Conv2d(32, 64, 3, padding=1), nn.BatchNorm2d(64), nn.ReLU(), nn.MaxPool2d(2),
            nn.AdaptiveAvgPool2d((1, 1))
        )
        self.classifier = nn.Linear(64, num_classes)
    
    def forward(self, x):
        x = self.features(x)
        x = x.flatten(1)
        return self.classifier(x)

Training uses cross-entropy loss and Adam optimizer with early stopping. For cross-room evaluation, the model is trained on 5 rooms and tested on the held-out room.

Mathematical Formulation§

For a given input representation $X \in \mathbb{R}^{T \times F}$ (time frames × frequency/range bins), the CNN learns a mapping $f_\theta: X \mapsto y$, where $y$ is a probability vector. Training minimizes:

$$ \mathcal{L}_{CE} = -\frac{1}{N}\sum_{i=1}^{N} \sum_{c=1}^{C} \mathbb{1}[y_i = c] \log \hat{p}_{i,c}, $$

where $\hat{p}_{i,c}$ is the predicted probability for class $c$.

To quantify the performance-robustness trade-off, the authors compute the difference between cross-subject and cross-room F1 scores:

$$ \Delta = F1_{cross\text{-}subject} - F1_{cross\text{-}room}. $$

IR-UWB has the largest positive $\Delta$ (i.e., high performance, lower robustness), while FMCW has near-zero or negative $\Delta$ (slightly better generalization).

Practical Guidelines§

• For fine-grained activity recognition in a fixed room, IR-UWB is preferred.
• For unseen environments or deployment across multiple rooms, FMCW is more reliable.
• For coarse sleep monitoring, all technologies are viable, making cost and hardware simplicity the deciding factors.
• Ceiling-mounted placement works effectively for all three, validating its practicality in healthcare settings.

Conclusion§

This is the first fair, controlled comparison of ceiling-mounted FMCW, IR-UWB, and Wi-Fi for healthcare sensing. The identified trade-off provides a principled basis for selecting RF technology based on deployment constraints and monitoring requirements.