Privacy-Preserving Radar Occupancy Sensing

• Privacy-preserving radar occupancy sensing is a method that utilizes mmWave FMCW radar to detect indoor occupancy through micro-Doppler and range signatures while ensuring privacy.
• The system integrates robust hardware, advanced signal processing pipelines, and machine learning algorithms, including hybrid quantum-classical approaches, to achieve high accuracy and real-time performance.
• Empirical results demonstrate 90–95% detection accuracy with sub-100 ms latency, meeting GDPR requirements and making it suitable for smart building management and elder care applications.

Privacy-preserving radar occupancy sensing refers to the use of radar—especially millimeter-wave frequency modulated continuous wave (mmWave FMCW) systems—as a means of monitoring occupancy in indoor environments without collecting personally identifying information (PII). Unlike video or RF tag-based methods, mmWave radar captures only physical micro-doppler and range–Doppler signatures, providing effective sensing of human presence, location, and motion while inherently preserving privacy. This modality suits applications such as smart building resource management and remote monitoring in elder care, where minimizing privacy risks is critical (Barral et al., 2024, Ratto et al., 17 Jan 2026).

1. mmWave FMCW Radar Architecture for Occupancy Sensing

mmWave FMCW radar modules operate by emitting a linearly frequency-modulated "chirp" and mixing the returned echo with the transmitted signal to obtain a beat frequency, enabling the estimation of range and radial velocity of targets. Key hardware parameters in recent research include:

• Chipset: TI IWR6843 (60–64 GHz), ARM R4F + DSP on chip
• Evaluation boards: IWR6843ISK (PCB antenna, elevation FoV ≈ 30°), IWR6843AOP EVM (AOP, elevation FoV ≈ 120°)
• RF settings: $f_c = 62\,\mathrm{GHz}$, $B = 4\,\mathrm{GHz}$, $\Delta R = c/(2B) \approx 0.0375\,\mathrm{m}$, $T_{chirp} = 60\,\mu\mathrm{s}$, $N_r = 256$, $f_s \approx 2\,\mathrm{MHz}$
• Doppler setup: $N_{chirps} = 128$/frame, $\Delta v = c/(2 f_c N_{chirps} T_{chirp})$
• Antenna deployment: Multiple radars (e.g., 2 wall-mounted and 1 ceiling-mounted) with measured position and orientation, using 3D-printed mounts, collectively covering a $12 \times 6\,\mathrm{m} = 72\,\mathrm{m}^2$ area with overlapping fields of view to mitigate occlusions (Barral et al., 2024).

The signal model is

$$s_{tx}(t) = \exp\left[j(2\pi f_c t + \pi S t^2)\right],$$

with the beat frequency $f_b = S\tau = S\frac{2R}{c}$ yielding range $R = \frac{c f_b}{2S}$, and Doppler shift $f_d = \frac{2 v_r f_c}{c}$ implying $v_r = \frac{c f_d}{2 f_c}$.

2. Signal Processing Pipelines and Algorithms

Low-level signal processing comprises:

• ADC sampling: $N_r=256$ at $f_s\approx 2\,\mathrm{MHz}$
• Range FFT: Windowed fast-time FFT per chirp to obtain range bins
• Doppler processing: Stacked slow-time frames subjected to FFT to produce range–Doppler maps
• Detection: CFAR on range–Doppler map, target $P_{FA}\approx 10^{-6}$
• Point cloud extraction: (range, Doppler, azimuth, elevation, SNR)

Filtering is applied via:

• Threshold filter: SNR and velocity gating
• Buffer filter: Sliding-window validation ($\Delta t_{buf}\approx 0.5\,\mathrm{s}$) to remove transient "ghosts" (Barral et al., 2024).

Spatial registration is performed using ROS TF2, transforming radar-frame detections into a global building frame before merging all radars' outputs into a single topic. Clustering algorithms such as DBSCAN (uniform density) and OPTICS (variable density) are applied to the fused point cloud to separate individual objects per time-slice ($\Delta t_{cluster}\approx0.1\,\mathrm{s}$). Target tracking uses an extended Kalman filter (EKF) with four-dimensional state ([X, Y, Vx, Vy]), and data association relies on Mahalanobis gating.

Occupancy is inferred either by counting tracked IDs/clusters in defined rectangular zones or by discretizing space into uniform cells, each with hysteresis timers ($T_{enter}$, $T_{leave}$) to determine long-term occupancy (Barral et al., 2024).

3. Software Infrastructure and IoT Integration

A modular, open-source ROS-based software stack supports the architecture:

• Acquisition layer: mmWave Reader nodes (USB interfacing, data packetization)
• Filtering layer: Threshold and buffer filters per radar stream
• Fusion layer: MultiRadarToSingleTopic merges and spatially registers outputs
• Clustering layer: mmWaveCluster delivers real-time clustering
• Tracking layer: TargetFollower maintains EKFs for track management
• Application layer: TimeOccupancyGrid outputs zone counts/events

All nodes communicate with ROS messages; occupancy summaries are exported using MQTT in JSON format. Time synchronization is enforced via NTP across devices. The system can be linked with Home Assistant by subscribing to MQTT topics, and occupancy is visualized via HA dashboards (Barral et al., 2024).

End-to-end latency is sub-100 ms (typical ≈ 80 ms), and per-radar chain CPU utilization on Raspberry Pi 4 is ~35%. Software is available: github.com/GTEC-UDC/mmwave_reader, mmwave_fuse, mmwave_cluster, door_counter.

4. Machine Learning and Digital Twin–Driven Privacy-Preserving Classification

Physics-informed digital twins are used to generate realistic synthetic range–Doppler map (RDM) data. Simulation relies on GO/PO hybrid shooting-and-bouncing-rays (SBR) solvers to model the electromagnetic backscattering from animated 3D occupancy scenarios, with explicit modeling of:

$$\mathbf{E}^i(\mathbf{r}) = \left(-\hat\phi^i I + \hat\theta^i\bar{I}\right)e^{j\mathbf{k}^i\cdot\mathbf{r}}$$

The radar aperture induces magnetic currents and yields simulated received signals allowing for detailed control over noise and environment (Ratto et al., 17 Jan 2026).

A hybrid quantum–classical neural network (HQNN) is benchmarked for occupancy classification (distinguishing 0, 1, or 2 occupants):

• Input: $128\times128$ RDM
• Classical frontend: Shallow CNN (1→16→32 channels), pooled to $2\times15\times15$, flattened, dense layers to $2$-D latent vector
• Quantum layer: Two-qubit parameterized quantum circuit (PQC): ZZFeatureMap, RealAmplitudes ansatz with 4 angles, entangling gates
• Measurement: Pauli $ZI$ and $IZ$ expectation values $(f_1, f_2)$, followed by a linear layer and softmax
• Parameter count: HQNN ≈ $0.066$M vs. DopplerNet $2.6$M, EfficientNet-B0 $4.0$M, ResNet-18 $11.2$M

HQNN achieves 99.7% accuracy on synthetic and 97.0% on real data. Classical controls without the PQC collapse to 68.5% and 31.5% balanced accuracy (BA) on real data. Under additive AWGN, HQNN recovers earlier in synthetic domains but CNNs recover more fully on real data at high SNR ($+$10 dB, HQNN BA = $0.842$, ResNet-18 BA = $0.984$). At 50% labeled data, HQNN BA drops to $0.75$ while EfficientNet-B0 maintains $0.99$ (Ratto et al., 17 Jan 2026).

5. Empirical Performance and Privacy Guarantees

Experimental deployment in a 72 $\,\mathrm{m}^2$ lab with 3 radars and up to 4 persons yields:

• Position accuracy: mean error ±0.3 m compared to camera ground truth
• Detection accuracy: 90–95% for moving occupants
• False positive rate: ~5%
• False negative rate: ~10% (mostly missed static occupants exceeding $T_{timeout}$)
• Fusion latency: ~80 ms
• Edge CPU load: ~35% per radar chain on RPi4 (Barral et al., 2024)

mmWave radar systems inherently preserve privacy:

• Only produce low-dimensional point clouds (range, velocity vectors) with no visual or biometric data
• No tags, no face or behavior identification, and no user cooperation required
• Satisfy GDPR-compliant privacy requirements: no visual or physical PII, minimum risk from data interception

Such systems provide privacy guarantees not possible with RGB imaging or user-carried devices, supporting use in sensitive environments (Barral et al., 2024, Ratto et al., 17 Jan 2026).

6. Trade-Offs, Design Guidelines, and Future Opportunities

Increasing the number of radars reduces occlusions and blind spots but at the cost of higher computational requirements and possible RF interference. Real-time rates (~10 Hz, $128$ chirps @ $60\,\mu$ s) are feasible; reducing chirp numbers decreases Doppler resolution but lowers latency and CPU load. DBSCAN clustering is efficient under uniform target density, while OPTICS handles variable density at approximately double the computational expense. Detecting static occupants requires radar configurations sensitive to zero-Doppler signatures or reduced CFAR thresholds.

On the machine learning front, quantum–classical networks provide competitive accuracy in highly parameter-efficient forms and exhibit distinctive robustness to moderate SNR in synthetic domains. However, on real, label-scarce data, classical CNNs retain sample efficiency advantages. The HQNN's memory and computational footprint makes on-device edge inference feasible, further strengthening privacy by limiting data egress (Barral et al., 2024, Ratto et al., 17 Jan 2026).

Continued development of physics-informed simulation (digital twins), advanced filtering/tracking, and parameter-efficient learning architectures will support the deployment of scaled, privacy-preserving radar occupancy sensors in diverse smart environments.