Massive Wireless Human Sensing

• Massive wireless human sensing is a paradigm that exploits heterogeneous, distributed RF signals to non-intrusively extract human presence, activities, and vital signs.
• It integrates device-free and device-based approaches using commodity infrastructures like Wi-Fi, radar, and multi-antenna systems to achieve robust, scalable monitoring.
• Recent methods combine multi-domain diversity, advanced signal processing, and scalable architectures to improve accuracy, reduce false alarms, and address deployment challenges.

Searching arXiv for papers on massive wireless human sensing and closely related systems. Massive wireless human sensing denotes a sensing paradigm in which a monitored area is instrumented not by a single specialized sensor, but by a heterogeneous, distributed, and opportunistically exploited wireless infrastructure that extracts human-related information from communication signals and sensing-oriented RF measurements. In the position-paper formulation of the field, the objective is to exploit diversity in the time, frequency, and space domains while combining device-free and device-based sensing approaches, so as to improve accuracy and service availability across indoor and outdoor environments (Sanctis, 13 Aug 2025). Across the literature, this umbrella encompasses commodity Wi-Fi sensing, Integrated Sensing and Communications (ISAC), massive MIMO and large intelligent surface architectures, mmWave and ultra-wideband radar, reconfigurable intelligent surfaces, and standardized data/benchmark protocols (Hu et al., 2023, Sakamoto, 2020, Wu et al., 18 Jul 2025, Manoj et al., 2021, Nelson et al., 2023, Li et al., 2024, Huang et al., 13 Dec 2025).

1. Conceptual scope and defining characteristics

The clearest explicit definition appears in the architectural formulation of Massive Wireless Human Sensing (MaWiS), which describes an infrastructure that opportunistically analyzes a plethora of heterogeneous wireless communication signals to extract physical information about people, including presence, occupancy or crowd counting, activity recognition, fall detection, localization and trajectory tracking, vital signs, gestures, and gait (Sanctis, 13 Aug 2025). In that formulation, “massive” refers simultaneously to frequency diversity across multiple bands and bandwidths, time diversity across packet types and schedules, space diversity across multiple nodes and multi-antenna receivers, multi-technology diversity, and the joint use of device-free and device-based sensing.

A closely related network-centric interpretation emerges in ISAC. There, sensing is embedded into communication infrastructure such that spectrum, time, and hardware are shared between connectivity and perception, with the resulting system reusing downlink and uplink signaling, massive MIMO arrays, multibeam analog beamforming, and distributed intelligence to support scalable situational awareness (Wu et al., 18 Jul 2025). This formulation treats “massive” not only as a statement about the number of users or sensors, but also about continuous operation, multi-cell cooperation, and physically interpretable feature pipelines that remain usable across heterogeneous deployments.

At the implementation level, the term also covers large-scale commercial deployment. A notable example is a WiFi sensing system for ubiquitous home monitoring deployed across over 10 million commodity routers and more than 100 million smart bulbs, demonstrating that scale can be realized through broad COTS adoption rather than only through laboratory array size or bandwidth (Zhu et al., 4 Jun 2025). This suggests that massive wireless human sensing is simultaneously an architectural, algorithmic, and operational notion: it concerns infrastructure scale, signal diversity, and the practical ability to sense many people, homes, rooms, or devices under realistic communication constraints.

A recurring distinction within this scope is between device-free and device-based modes. Device-free sensing interprets perturbations that human bodies imprint on ambient RF propagation, whereas device-based sensing passively analyzes transmissions from mobile devices carried by people without requiring tags, apps, or explicit collaboration (Sanctis, 13 Aug 2025). This duality is central because many large-scale systems combine both regimes rather than treating them as mutually exclusive.

2. Physical foundations and separability mechanisms

Across the literature, human sensing is grounded in the fact that motion, posture, physiology, and environmental interaction perturb wireless propagation through reflection, diffraction, scattering, delay variation, Doppler, and angle dispersion. The MaWiS formulation expresses core primitives through channel state information $H_k(n) = |H_k(n)| e^{j\phi_k(n)}$, time-varying impulse responses $$h(\tau,t)=\sum_i \alpha_i(t)\,\delta(\tau-\tau_i(t))$$, angle relations such as $\Delta\phi = \frac{2\pi d}{\lambda}\sin\theta$, and monostatic Doppler $f_D=\frac{2vf_c}{c}$ (Sanctis, 13 Aug 2025). ISAC reviews extend this to a multi-antenna OFDM CFR model in which delays, Doppler shifts, angle of arrival, angle of departure, CFO, and hardware-induced phase offsets all coexist in the measured channel (Wu et al., 18 Jul 2025).

In radar-centric sensing, separability is often obtained directly from bandwidth and aperture. Ultra-wideband radar leverages high range resolution, array processing, and frequency–wavenumber or reversible-transform imaging to separate multiple human scatterers, while vital-sign sensing exploits phase modulation induced by chest displacement in baseband models such as $s(t)=A e^{j2k d(t)} + s_{DC}$ (Sakamoto, 2020). In that regime, large bandwidth is the primary lever for disambiguating multiple targets.

Commodity Wi-Fi sensing confronts a different separability problem because limited bandwidth yields coarse range resolution. The near-field Wi-Fi approach of MUSE-Fi addresses this by exploiting physical dominance rather than classical resolution: when a Wi-Fi device is very close to a subject, near-field channel variation caused by that subject significantly overwhelms variations caused by distant subjects, so CSI carried by traffic to and from that device becomes subject-specific (Hu et al., 2023). The paper frames this as physical separability and instantiates it through uplink CSI, downlink CSI, and downlink beamforming feedback. A plausible implication is that large-scale multi-person sensing on commodity infrastructure may depend less on resolving every person from a single link than on constructing many locally dominant links.

Other works realize separability through spatial aperture. Massive MIMO sensing constructs third-order channel tensors over time, frequency, and space, then extracts low-rank factors from correlations across these domains; with a 100-antenna array, activity-classification accuracy reached 98% in LoS and 87% in NLoS, substantially above smaller-array baselines (Manoj et al., 2021). Large Intelligent Surface and RadioWeave measurements at 5.6 GHz further show strong spatial non-stationarity, with up to 60 dB variation across a very large distributed aperture, implying that different sub-arrays see different visibility regions and can therefore isolate localized human-induced changes (Nelson et al., 2023).

Reconfigurable environments introduce yet another separability mechanism. Space-time-coding RISs generate frequency-orthogonal harmonic beams, assigning distinct harmonics to different detected persons and thereby separating respiration and heartbeat channels in both frequency and space (Li et al., 2024). This is not equivalent to bandwidth-based ranging; rather, it is programmable physical-layer multiplexing. Taken together, the literature shows that massive sensing does not rely on a single separability principle. It may emerge from UWB resolution, large-aperture angular discrimination, near-field dominance, harmonic multiplexing, or distributed multi-link diversity, depending on hardware and deployment assumptions.

3. System architectures and wireless modalities

The MaWiS architecture centers on the “massive wireless human sensing edge device,” a receive-only, passive embedded node integrating multiple wireless front-ends, multi-antenna configurations, baseband capture, protocol-aware parsers, timestamping, on-device signal conditioning, and secure feature upload to servers for inference and training (Sanctis, 13 Aug 2025). The cited implementation path includes commodity edge compute such as Raspberry Pi, NVIDIA Jetson, or Intel NUC, paired with WiFi NICs, Bluetooth radios, LoRa modules, 4G/5G receivers, DVB/DAB SDRs, and general SDRs.

Different sensing modalities contribute complementary observables. WiFi contributes CSI and CIR with rich subcarrier structure; Bluetooth and LoRa contribute lower-rate but prevalent or long-range packet streams; 4G/5G/6G contribute wide coverage, multi-band operation, and beam or reference-signal metrics; DVB-T and DAB offer continuous waveforms for passive sensing; mmWave, FMCW, and UWB radar contribute micro-Doppler, fine range, and angle estimation (Sanctis, 13 Aug 2025). The multimodal dataset MM-Fi concretizes this heterogeneity by synchronizing RGB, depth, LiDAR, mmWave radar, and WiFi CSI into a common 10 Hz frame rate across 320.76k frames from 40 subjects and 27 actions (Yang et al., 2023). Although that dataset includes vision modalities, its wireless components—LiDAR, mmWave radar, and WiFi—illustrate the breadth of non-intrusive sensing channels now considered jointly.

ISAC broadens the architectural space further by embedding sensing directly into standardized communication infrastructure. In the 802.11ay RAPID system, 60 GHz access points reuse beam training and tracking fields to estimate CIRs and extract micro-Doppler without modifying packet structure, enabling people tracking, HAR, and person identification from standard-compliant phased-array WLAN equipment (Pegoraro et al., 2021). In the Wi-Fi 7 people-counting test-bed, pattern-reconfigurable antenna systems provide beam-space diversity with beam codes 21 and 42, while split learning keeps raw CSI local and sends only local decisions for aggregation (Bersan et al., 15 Jun 2026). These examples show two distinct scaling paths: sensing-friendly standard features in high-frequency WLANs, and edge ML architectures that fit within commodity access-point constraints.

Distributed infrastructure also appears in large-aperture sub-6 GHz concepts such as RadioWeaves. There, coherent processing across many embedded radio elements and distributed internal compute resources enables multistatic geometries, local feature extraction, and hierarchical fusion across walls or ceilings (Nelson et al., 2023). This suggests that “massive” can refer not only to many users or packets but also to physically large, deeply integrated sensing surfaces.

4. Signal processing, inference, and data representations

Despite hardware diversity, the sensing pipeline in the literature has a recurrent structure: acquisition, synchronization or calibration, clutter suppression, feature extraction, fusion, and task-specific inference. The MaWiS pipeline states this explicitly, listing phase calibration, clock-drift compensation, antenna calibration, resampling and interpolation, denoising such as SVD-based filtering and wavelet denoising, and feature extraction spanning CSI statistics, differential phase, Doppler spectrograms, power variance, and packet-level timing features (Sanctis, 13 Aug 2025). It also formalizes fusion through Bayesian inference, weighted least squares, Kalman filtering, and multi-modal deep learning.

Several works emphasize preprocessing that is robust to heterogeneous commodity hardware. The millions-scale home-monitoring deployment deliberately adopts amplitude-only processing to avoid phase errors from CFO, SFO, and IQ imbalance, computing $G(t,f)=|H(t,f)|^2$, an autocorrelation function $\rho_G(\tau,f)$, a motion statistic $\phi(f)=\rho_G(\Delta t,f)$, and maximal ratio combining across subcarriers to improve SNR by 8–12 dB in practice (Zhu et al., 4 Jun 2025). Speed is estimated from the first peak of the ACF-derived statistic via the Bessel-function relation reported in that work, and human versus non-human recognition uses gait-consistent ACF features with an SVM. This is a particularly strong example of signal processing chosen for deployment robustness rather than maximal expressiveness.

At the opposite end of the model spectrum, generative-AI-assisted sensing uses a unified weighted conditional diffusion model to denoise velocity–acceleration spectra and resolve ambiguous DoA spectra in downlink Wi-Fi human-flow detection (Wang et al., 2024). There, CSI from a reference antenna and a surveillance ULA is transformed through conjugate multiplication, DC nulling, a symmetric instantaneous autocorrelation, a keystone transform, and 2D FFT to produce a velocity–acceleration spectrum. Conditional diffusion then suppresses cross terms and noise, while separate 2D-MUSIC stages estimate DoA and ToF, followed by clustering into subflows and subflow sizes. The reported subflow-size detection accuracy reaches 91% (Wang et al., 2024).

Tensor methods form another recurring thread. Massive MIMO activity sensing organizes channels into tensors $\mathscr{G}\in\mathbb{C}^{T\times F\times M}$, derives time–frequency–space correlation tensors, applies CP decomposition via ALS, and feeds concatenated eigenvalue features into a compact neural classifier (Manoj et al., 2021). The later Sensing Dataset Protocol generalizes this idea into a protocol-level representation, mapping heterogeneous measurements into a canonical tensor $X\in\mathbb{C}^{A\times K\times T}$, optionally using real–imag or amplitude–phase stacking, and then applying CP–ALS pooling to preserve spatial, spectral, and temporal structure in a task-agnostic descriptor (Huang et al., 13 Dec 2025). In the benchmark reported there, cross-user evaluation shows approximately 88% variance reduction across seeds while maintaining competitive accuracy and low latency (Huang et al., 13 Dec 2025). This suggests that representation standardization is becoming part of the core technical stack for massive sensing, not merely an evaluation convenience.

Location-aware decomposition is also central in ISAC pipelines. The unified ISAC review highlights sequential Doppler FFT, delay-domain MVDR, AoA FFT, CFAR detection, EKF tracking, and location-based extraction of residual CSI phase for respiration and heartbeat (Wu et al., 18 Jul 2025). In RAPID, per-person micro-Doppler spectrograms are obtained by selecting the beam and delay bins associated with each tracked subject, then feeding those spectrograms to residual CNNs for HAR and identification (Pegoraro et al., 2021). A common pattern thus emerges: scalable sensing increasingly depends on physically structured intermediate representations—Doppler–Delay–AoA cubes, point clouds, tensors, or compact ACF descriptors—rather than on raw signal ingestion alone.

5. Representative applications and empirical performance

The application range documented in the cited papers is broad. Massive sensing targets presence detection, occupancy estimation, people counting, activity recognition, trajectory tracking, gesture detection, respiration, heartbeat, heart-rate variability, and person identification (Sanctis, 13 Aug 2025, Sakamoto, 2020). The most comprehensive empirical picture comes from combining architectural papers, large-scale deployments, and modality-specific test-beds.

In commercial-scale WiFi home monitoring, a system operating across over 10 million routers and more than 100 million smart bulbs achieved 92.61% human motion detection accuracy across two years of testing, while reducing false alarms due to non-human movements from 63.1% to 8.4% and reducing cloud transmission overhead by 99.72% (Zhu et al., 4 Jun 2025). The evaluation involved 280 edge devices, 16 scenarios, and over 4 million motion samples. The same work reports human-versus-non-human recognition accuracy increasing from 37.9% to 90.4% across 7 homes, 84 days, and 2.4 million labeled samples (Zhu et al., 4 Jun 2025). These figures are noteworthy because they come from heterogeneous, low-cost chipsets and coexist with ordinary 802.11n/ac/ax traffic.

In massive-flow sensing using commodity downlink Wi-Fi, target-counting detection accuracy reached 92% in a corridor under download traffic, subflow-number detection reached 93%, and subflow-size detection reached 91% (Wang et al., 2024). The same work reports median DoA errors around 6 degrees when resolving grating-lobe ambiguity at $$h(\tau,t)=\sum_i \alpha_i(t)\,\delta(\tau-\tau_i(t))$$0 באמצעות conditional diffusion, close to the $$h(\tau,t)=\sum_i \alpha_i(t)\,\delta(\tau-\tau_i(t))$$1 baseline. This indicates that generative denoising can compensate for hardware configurations that would otherwise be unsuitable for fine angular sensing.

At higher frequencies, RAPID reports HAR accuracy of 94% and person-identification accuracy of 90% using two IEEE 802.11ay access points and CIR-derived micro-Doppler signatures (Pegoraro et al., 2021). The 1.76 GHz bandwidth gives a nominal range resolution of approximately 8.5 cm, enabling reliable multi-subject tracking with standard-compliant 60 GHz APs. In the radar domain, the UWB review reports approximately 1% instantaneous heart-rate error and an interbeat-interval RMSE of 7.9 ms relative to ECG in one topology-based heart-rate method, with additional average interbeat-interval errors of 14.0 ms from the sole of the foot and 16.3 ms from the head using MIMO combining (Sakamoto, 2020).

RIS-assisted multiperson vital-sign monitoring provides another empirical point. A space-time-coding RIS prototype simultaneously monitored the vital signs of up to four persons, with respiration-rate and heartbeat-rate estimation errors below 1 RPM and 5 BPM, respectively, when using the full improved VMD algorithm in the reported summary (Li et al., 2024). The same paper attributes the result to flexible beam control that reduces noise from irrelevant objects and improves the SNR of echoes from the human chest.

Dataset-based studies reveal modality trade-offs under controlled benchmarking. In MM-Fi, under random split and all 27 actions, LiDAR achieved MPJPE $$h(\tau,t)=\sum_i \alpha_i(t)\,\delta(\tau-\tau_i(t))$$2 mm, mmWave achieved PA-MPJPE $$h(\tau,t)=\sum_i \alpha_i(t)\,\delta(\tau-\tau_i(t))$$3 mm, and WiFi achieved MPJPE $$h(\tau,t)=\sum_i \alpha_i(t)\,\delta(\tau-\tau_i(t))$$4 mm for 3D pose estimation; under cross-environment evaluation, mmWave was the most robust wireless modality with MPJPE $$h(\tau,t)=\sum_i \alpha_i(t)\,\delta(\tau-\tau_i(t))$$5 mm (Yang et al., 2023). Wireless-only late fusion of radar, LiDAR, and WiFi achieved PA-MPJPE 42.7 mm under the random split, showing that privacy-preserving multimodal fusion can outperform individual modalities (Yang et al., 2023).

These results do not establish a single dominant technology. Rather, they show that the performance frontier depends strongly on task, modality, environment shift, communication constraints, and whether scale is defined by number of people, number of devices, or number of deployed nodes.

6. Scalability challenges, standardization, and research directions

Two persistent technical bottlenecks recur across the literature: heterogeneous data fusion and training generalization. The MaWiS position paper identifies specialized heterogeneous data fusion and feasible training phases as the two central challenges, emphasizing alignment of asynchronous measurements, missing-data handling, early versus late fusion, and the brittleness of environment-specific training (Sanctis, 13 Aug 2025). ISAC reviews add robust phase and asynchrony removal, multi-target separation under low bandwidth, generalizable feature abstractions, and calibration against environmental confounders as core unresolved issues (Wu et al., 18 Jul 2025).

Large-scale deployment studies reveal additional operational limits. In ubiquitous home monitoring, motion interference in multi-user environments remains difficult when occupants are close to the same device, CSI quality varies across chipsets and placements, and even lightweight edge classifiers can overload consumer routers at modest sensing rates (Zhu et al., 4 Jun 2025). In the Wi-Fi people-counting test-bed, static occupancy is less salient than motion-induced changes, beampattern sensitivity is unequal across beam codes, and spectrum conditions outside the pristine test house may degrade reliability (Bersan et al., 15 Jun 2026). These observations indicate that the transition from lab-scale inference to massive deployment is constrained as much by packet cadence, MAC coexistence, and hardware variability as by raw sensing accuracy.

Standardization and reproducibility have therefore become first-order concerns. The Sensing Dataset Protocol proposes a unified data-block schema, deterministic ordering, fixed CP rank and ALS sweeps, cross-user evaluation, and consistent reporting, explicitly targeting the fragmentation that impedes fair comparison and transfer across datasets and modalities (Huang et al., 13 Dec 2025). This protocol-level move is significant because massive sensing requires not only many sensors, but also stable abstractions through which heterogeneous measurements can be compared, fused, and benchmarked.

Infrastructure evolution is also reshaping the field. Wi-Fi 7 and WLAN sensing standardization provide wider bandwidths and more usable pilot subcarriers, while 802.11ay and future 802.11bf introduce sensing-friendly beam training, beam tracking, and likely API exposure (Pegoraro et al., 2021, Bersan et al., 15 Jun 2026). ISAC roadmaps point to massive MIMO, multibeam analog beamforming, STAR antennas, programmable metasurfaces, 4D arrays, and distributed semantic sensing as enabling substrates for scalable human and environmental sensing (Wu et al., 18 Jul 2025). Large distributed apertures such as RadioWeaves suggest a path in which the built environment itself becomes a sensing surface (Nelson et al., 2023).

A plausible implication is that future massive wireless human sensing will be layered rather than monolithic. Wide-coverage, low-cost sub-6 GHz systems may provide persistent coarse sensing and occupancy analytics; mmWave and UWB systems may supply high-resolution localization or vital-sign estimation in selected zones; RISs and large intelligent surfaces may improve physical-layer separability; and unified protocols such as SDP may provide common machine-learning interfaces across these modalities (Huang et al., 13 Dec 2025). The field’s trajectory, as reflected in the cited papers, is therefore toward pervasive but heterogeneous sensing fabrics in which scale is achieved by coordination, complementarity, and reproducible abstractions rather than by any single waveform or device class.