📡 Wireless Sensing — A Primer

Why this topic? Your smartphone's WiFi isn't just connecting you to the internet — it's silently painting a picture of everyone moving around you. Wireless sensing turns everyday radio signals into invisible motion detectors, no cameras or wearables required.

Background

Imagine dropping a pebble into a still pond. The ripples spread outward in perfect, predictable circles — until they hit something. A rock, a reed, your hand. The pattern distorts. Now imagine the pond is invisible radio waves bouncing between a WiFi router and a receiver, and the "pebble" is a person walking across the room. Every step they take disturbs those waves in a measurable way. That's the core intuition behind wireless sensing.

WiFi routers constantly broadcast radio signals. Normally, we only care whether those signals delivered our data. But the signal itself carries a fingerprint of everything it passed through or bounced off. Researchers noticed that this fingerprint changes whenever something in the environment moves — a person walking, breathing, even their heartbeat. Rather than throwing away this rich physical information, wireless sensing systems harvest it.

The dominant signal metric used today is Channel State Information (CSI) — a detailed measurement of how a radio wave changes as it travels from transmitter to receiver. Unlike the cruder Received Signal Strength Indicator (RSSI), which just measures total signal power (imagine only knowing how loud a sound is, not its pitch), CSI captures amplitude and phase across many individual frequency channels simultaneously. This richness makes it far more sensitive to subtle environmental changes. The foundational tool enabling this on commodity hardware was released by halperin2011_2cbc, which demonstrated that standard 802.11n WiFi cards could record detailed CSI alongside normal packet captures — opening the field to academic researchers worldwide.

The practical motivation is compelling. Camera-based systems raise privacy concerns and fail in low-light conditions. Wearable sensors require users to opt in and charge devices. Dedicated radar hardware is expensive. WiFi infrastructure, by contrast, already blankets offices, hospitals, transit hubs, and homes. Repurposing it for sensing costs almost nothing extra and works through walls, in the dark, without anyone's active participation.

Key Methods

The field has converged on a pipeline of methods, each solving a different part of the problem:

Signal extraction uses CSI Fingerprinting to capture how specific locations or activities produce repeatable CSI patterns. The raw signal travels across many subcarriers — individual frequency slices within a WiFi channel that OFDM (Orthogonal Frequency-Division Multiplexing, the technology that lets WiFi pack data onto many frequencies simultaneously) separates out — giving researchers tens to hundreds of parallel observations per packet.

Preprocessing typically involves Principal Component Analysis to strip out static background reflections and isolate the dynamic components caused by motion, alongside Fourier Transforms to convert time-domain signal variation into frequency-domain features (e.g., detecting breathing rate as a regular oscillation).

Inference is dominated by deep learning: Convolutional Neural Networks treat CSI snapshots like images, Long Short-Term Memory networks model temporal sequences of signal change, and Support Vector Machines remain competitive for lower-data regimes. The challenge is generalisation — a model trained in one room often fails catastrophically in another.

Active Research in the Vault

The survey ma2020_4782 established the canonical taxonomy of CSI-based applications — localization, activity recognition, gesture recognition, and vital sign detection — and remains the standard entry point for newcomers. Building on this foundation, wang2026_2758 systematically reviews over 200 papers to identify why systems fail when moved to new users or environments, proposing a four-stage pipeline for diagnosing generalisation failures.

For people counting specifically, golammowla2024_1b1f demonstrates a CNN+LSTM model counting 1–5 persons with commodity hardware, while torun2026_72aa tackles the harder problem of seated (stationary) crowds by analysing subtle body-fidget signal bandwidth rather than gross motion. choi2022_17c2 extends this further to simultaneously count and localise crowds within a room.

Reproducibility is a growing concern. guarino2026_e72c audits the literature and finds most results are not independently verifiable, while meneghello2023_0a93 directly addresses this by releasing a richly annotated public dataset spanning 13 subjects and 7 environments. The open-source taehyeon2026_4bd5 framework built on commodity ESP32 microcontrollers further lowers the barrier to entry.

Environment dependence — the tendency of models to overfit to a specific room's geometry — is addressed by jiang2018_77f6, which uses adversarial training to suppress environment-specific features, and santos2024_1e39, which empirically quantifies how much accuracy degrades across 7 different rooms. zabin2026_a20c proposes a physics-informed approach to domain adaptation that selects subchannels based on their physical stability.

Beyond WiFi, the field uses multiple RF technologies. billah2021_69a2 shows Bluetooth Low Energy can detect occupancy with reinforcement learning that adapts as furniture moves. hu2024_01d0 uses millimetre-wave (mmWave) radar — operating at much higher frequencies than WiFi, giving finer spatial resolution — for stationary crowd counting. pronello2025_9dc1 takes a different angle, using WiFi probe requests (the "hello, is anyone there?" messages devices broadcast when searching for networks) rather than CSI, enabling outdoor deployment without infrastructure access.

Energy efficiency has emerged as a parallel concern. koo2026_a08d surveys the trade-offs between model accuracy and power consumption, while sharma2024_c8a2 proposes opportunistic sensing that piggybacks on existing network traffic instead of generating dedicated sensing packets.

The standardisation of WiFi sensing is underway: ropitault2024_d49d describes the 802.11bf amendment, which will eventually make sensing a native, interoperable feature of all WiFi hardware — a potential inflection point for the entire field. Meanwhile, aljarrah2023_e060 and zhou2022_09d5 examine ISAC (Integrated Sensing and Communications) — next-generation systems where sensing and data transmission share the same waveform simultaneously.

For activity recognition specifically, yang2023_a34a provides a unified benchmark across multiple datasets, making it easier to compare competing deep learning approaches fairly. hu2023_5891 solves the multi-person disambiguation problem by exploiting near-field effects when a device is held close to the body. huang2025_060d provides a purpose-built dataset for multi-user scenarios, showing WiFi models can exceed 90% accuracy in identifying which of several people is performing an activity.

Open Problems & Gaps

• How can CSI-based crowd counting models be made environment-agnostic without per-room calibration? Current systems degrade sharply when deployed in new spaces (documented by santos2024_1e39); domain adaptation methods like those in zabin2026_a20c are promising but not yet production-ready.

• Can wireless sensing scale to truly dense crowds (50+ people) in complex environments? Nearly all crowd-counting benchmarks test fewer than 10 people in laboratory rooms; real transit or stadium scenarios involve hundreds of people, multi-path interference, and dynamic furniture — a regime almost entirely unexplored.

• How do we build trustworthy, reproducible baselines? guarino2026_e72c shows most published results cannot be independently verified; the community needs shared evaluation protocols and open datasets that cover diverse real-world conditions beyond the lab.

• What are the privacy boundaries of wireless sensing? The same capabilities that make WiFi sensing useful for crowd monitoring also enable covert surveillance; ficara2024_f89b highlights how MAC address randomisation attempts to protect users, but CSI-level sensing bypasses this entirely — raising unresolved ethical and regulatory questions.

History