| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| pretty_name: monostatic full-duplex Wi-Fi gesture sensing |
| --- |
| |
| # WIRD-GEST Dataset |
|
|
| Gesture recognition dataset collected via **monostatic full-duplex Wi-Fi sensing** on commercial off-the-shelf (COTS) laptops — no external sensors, no dedicated transmitter, no hardware modification of any kind. |
|
|
| Accompanying paper: **"WIRD-GEST: Gesture Recognition in the Real World Using Active Range-Doppler Wi-Fi Sensing on COTS Hardware"** (Sanson et al., 2025). |
|
|
| --- |
|
|
| ## Key Innovation: Monostatic Sensing |
|
|
| Most Wi-Fi sensing datasets use a **bistatic** setup: a separate transmitter (e.g., a router) and a receiver capture CSI between two devices. This requires coordinating two pieces of hardware and a line-of-sight path between them. |
|
|
| This dataset uses a **monostatic** setup instead. A single unmodified laptop simultaneously transmits and receives by sharing its Local Oscillator and baseband processing — the device's own self-interference becomes the sensing signal. CSI is read directly from the built-in NIC. No second device, no external transmitter, no hardware modification of any kind is required. |
|
|
| --- |
|
|
| ## Hardware & Capture Parameters |
|
|
| | Parameter | Value | |
| |---|---| |
| | Hardware | Lenovo ThinkPad (Wi-Fi 6E) | |
| | Bandwidth | 160 MHz | |
| | Frame rate | ~40 Hz | |
| | Channel | 79 (Fc ≈ 6.3 GHz) | |
| | Subcarriers | 512 (data subcarriers only, pilots removed) | |
| | LTF frames | 2 (csi1, csi2) — 1 RX antenna | |
|
|
| --- |
|
|
| ## Dataset Overview |
|
|
| | Property | Value | |
| |---|---| |
| | Participants | 5 (users 1–5) | |
| | Gesture classes | 5 | |
| | Session folders | 55 (50 lab + 5 café) | |
| | Total raw frames | 191,442 | |
| | Complete gesture instances | 722 | |
| | Collection environments | Lab (primary) + café/public space (cross-location subset) | |
| | Disk size | ~12 GB | |
|
|
| Sessions are pre-split into **train** (25 folders, 5 users × 5 gestures) and **val** (25 lab + 5 café = 30 folders) sets. The **café subset** (user 1 only) provides an out-of-environment evaluation split for cross-location generalisation. |
|
|
| --- |
|
|
| ## Gesture Classes |
|
|
| | Class key | Description | |
| |---|---| |
| | `hand_forward_back` | Forward / backward wave | |
| | `hand_up_down` | Up / down wave | |
| | `hand_pulse` | Pulse (push forward and back) | |
| | `hand_clock` | Clockwise circular motion | |
| | `hand_side` | Side-to-side wave | |
|
|
| --- |
|
|
| ## Folder Structure |
|
|
| ``` |
| gesture_wifi_monostatic_dataset/ |
| ├── dataset_metadata.json # aggregate statistics for the full dataset |
| ├── lenovo_user1_clock_train/ |
| │ ├── csi.csv |
| │ ├── metadata.yaml |
| │ └── range_doppler_data_32/ |
| │ └── range_doppler_frames_1.pkl |
| ├── lenovo_user1_clock_val/ |
| │ └── ... |
| ├── lenovo_user{1-5}_{clock,front,pulse,side,up}_{train,val}/ |
| │ └── ... # one folder per user × gesture × split (50 total) |
| └── cafe/ |
| └── lenovo_user1_public_space_{clock,front,pulse,side,up}_val/ |
| └── ... # cross-location evaluation subset (5 folders) |
| ``` |
|
|
| **Naming convention:** `lenovo_user<ID>_<gesture>_<split>` |
|
|
| Each session folder contains: |
| - `csi.csv` — calibrated CSI measurements with per-frame gesture labels |
| - `metadata.yaml` — session-level metadata (participant, gesture, date, frame statistics) |
| - `range_doppler_data_32/range_doppler_frames_1.pkl` — pre-processed range-Doppler frames |
|
|
| --- |
|
|
| ## Data Files |
|
|
| ### `csi.csv` — Raw CSI |
|
|
| Each row is one measurement frame (~25 ms interval at 40 Hz). |
|
|
| #### Calibration State |
|
|
| The CSI samples are frequency-domain measurements that have already been pre-processed: |
|
|
| - **Pilot subcarriers removed** — only the 512 data subcarriers are retained. |
| - **Phase and delay calibrated** — carrier frequency offset and timing offset compensation has been applied. |
| - **Two LTF frames averaged per measurement frame** — `csi1` and `csi2` columns hold the two averaged LTF measurements. |
|
|
| The data is ready for direct 2D DFT processing to produce range-Doppler maps. No additional calibration or pilot removal is required. |
|
|
| #### CSV Columns |
|
|
| | Column | Description | |
| |---|---| |
| | `event_timeStamp` | Device event timestamp (integer, ms) | |
| | `unix_timestamp` | Unix time in seconds (float) | |
| | `channel` | Wi-Fi channel number | |
| | `bandwidth_MHz` | Capture bandwidth in MHz | |
| | `measurement_time_repetition_ms` | Target frame interval in ms | |
| | `frequency_carrier_MHz` | Carrier frequency in MHz | |
| | `subcarrier_number` | Number of data subcarriers (512) | |
| | `csi1-{i}-real` | LTF 1, subcarrier i, real part (i = 0..511) | |
| | `csi1-{i}-imag` | LTF 1, subcarrier i, imaginary part (i = 0..511) | |
| | `csi2-{i}-real` | LTF 2, subcarrier i, real part (i = 0..511) | |
| | `csi2-{i}-imag` | LTF 2, subcarrier i, imaginary part (i = 0..511) | |
| | `label` | Gesture label: `front`, `up`, `pulse`, `clock`, or `side` | |
| | `start_end` | Gesture boundary marker: `start`, `end`, or `none` | |
|
|
| **`start_end` semantics:** Each gesture instance is bracketed by a `start` marker at the first frame of the motion and an `end` marker at the last frame. Frames outside any gesture instance are marked `none`. This allows precise extraction of complete gesture instances from the continuous recording. |
| |
| --- |
| |
| ### `range_doppler_data_32/range_doppler_frames_1.pkl` — Pre-processed Radar Frames |
| |
| A Python pickle file containing a **list of frame dictionaries**, one entry per measurement frame (~40 Hz), in temporal order. |
| |
| #### Per-frame dictionary keys |
| |
| | Key | Type | Description | |
| |---|---|---| |
| | `range_doppler_snr` | 2D numpy array | Range-Doppler heatmap, SNR values in dB | |
| | `timestamp` | int | Unix timestamp — matches `unix_timestamp` in `csi.csv` | |
| |
| #### Processing pipeline applied |
| |
| 1. The two LTF measurements (`csi1`, `csi2`) from `csi.csv` are averaged per frame. |
| 2. A 2D DFT is applied across the subcarrier (range) and time (Doppler) axes. |
| 3. Range and Doppler axes are interpolated to the cell sizes listed below. |
| 4. SNR is computed in dB and clipped to [5, 40] dB, then normalised to [0, 1]. |
| 5. The heatmap is resized to 64 × 64 and stored as `range_doppler_snr`. |
| |
| #### Range-Doppler map properties |
| |
| | Property | Value | |
| |---|---| |
| | Image size | 64 × 64 (range bins × Doppler bins) | |
| | Range axis | 0 to 0.63 m | |
| | Velocity axis | ±0.45 m/s | |
| | Range cell size | 0.93 cm | |
| | Doppler cell size | 0.015 m/s | |
| | Range resolution (physical) | 0.93 cm (from 160 MHz bandwidth) | |
| | Doppler resolution (physical) | 0.03 m/s (from 40 Hz frame rate) | |
| | Unambiguous velocity | ±0.4 m/s | |
| |
| --- |
| |
| ### `metadata.yaml` — Session Metadata |
| |
| A small YAML file present in every session folder summarising that session's recording. |
| |
| | Field | Description | |
| |---|---| |
| | `dataset.user` | Participant identifier (e.g. `user1`) | |
| | `dataset.gesture` | Gesture type for this session (e.g. `clock`) | |
| | `dataset.data_function` | Split: `training` or `validation` | |
| | `dataset.PC` | Recording machine identifier | |
| | `dataset.date` | Recording date (DD_MM_YYYY) | |
| | `csi.number_of_gestures` | Number of complete gesture instances in the session | |
| | `csi.frames_min` | Minimum frame count across gesture instances | |
| | `csi.frames_median` | Median frame count across gesture instances | |
| | `csi.frames_max` | Maximum frame count across gesture instances | |
| | `summary.number_of_gestures` | Total gesture instances (same as `csi.number_of_gestures`) | |
| |
| --- |
| |
| ### `dataset_metadata.json` — Aggregate Statistics |
| |
| A top-level JSON file with aggregate counts across all sessions. |
| |
| | Field | Description | |
| |---|---| |
| | `dataset_summary.total_samples` | Total raw frames across all sessions (191,442) | |
| | `dataset_summary.gesture_samples` | Frames labelled as a gesture (77,538) | |
| | `dataset_summary.none_samples` | Frames labelled as background / none (113,904) | |
| | `dataset_summary.train_samples` | Total frames in train sessions (125,898) | |
| | `dataset_summary.val_samples` | Total frames in val sessions (65,544) | |
| | `complete_gesture_breakdown` | Complete gesture instance counts per class | |
| | `folder_details` | Per-session sample and gesture counts | |
| |
| **Complete gesture instances per class:** |
|
|
| | Gesture | Instances | |
| |---|---| |
| | clock | 142 | |
| | front | 146 | |
| | pulse | 145 | |
| | side | 145 | |
| | up | 144 | |
| | **Total** | **722** | |
|
|
| --- |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{sanson2025wirdgest, |
| title = {WIRD-GEST: Gesture Recognition in the Real World Using Active |
| Range-Doppler Wi-Fi Sensing on COTS Hardware}, |
| author = {Sanson, Jessica Barthold and Shah, Rahul C. and Zhu, Yazhou |
| and Rosales, Rafael and Frascolla, Valerio}, |
| year = {2026}, |
| note = {IEEE ICC 2026 Workshop}, |
| } |
| ``` |
|
|
| --- |
