README.md

---
license: cc-by-nc-sa-4.0
pretty_name: RF-Behavior
language:
  - en
tags:
  - radar
  - mmWave
  - RFID
  - LoRa
  - IMU
  - motion-capture
  - human-activity-recognition
  - gesture-recognition
  - affective-computing
  - multimodal
task_categories:
  - time-series-forecasting
size_categories:
  - 10K<n<100K
viewer: false
extra_gated_prompt: >-
  RF-Behavior is released for academic research only, under CC BY-NC-SA 4.0.
  By requesting access you confirm that you will use the data for non-commercial
  research, that you will not try to identify the participants, and that you
  will cite the paper in work that uses the data.
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
  I agree to the terms above: checkbox
---

# RF-Behavior

A multimodal radio-frequency dataset for activity and affective behavior analysis.
61 participants (44 in the laboratory, 17 more in a living room and an
industrial site), plus 7 participants of a physiological reference subset,
performed **21 gestures**, **10 activities**, and **6 sentiment
expressions**, recorded at the same time by **13 mmWave radars** (8 on the
ground, 5 on the ceiling), **passive RFID tags** on the arms, a **LoRa link**,
**body-worn IMUs**, and **24 infrared motion-capture cameras**.

Paper: *RF-Behavior: A Multimodal Radio-Frequency Dataset for Activity and
Affective Behavior Analysis* — Si Zuo, Yuqing Song, Jin Han, Sahar Golipoor,
Ying Liu, Xujun Ma, Petter Holme, Stephan Sigg —
[arXiv:2511.06020](https://arxiv.org/abs/2511.06020). Project website:
https://sizuo.github.io/RF-Behavior/

The dataset has three environments: the **laboratory** (environment 1, all
five modalities), a **living room** (2) and an **industrial site** (3), the
latter two with the radars and the IMUs. The physiological reference subset
adds simultaneous ECG and EEG recordings in the living room.

## 1. Contents

| Campaign | Classes | Participants | Trials | Typical length |
|---|---|---|---|---|
| C1 gestures | 21 hand and arm gestures | 25 | 4,193 | 3.5 s |
| C2 activities | 10 activities (walking, running, sitting, lying, stairs, ball sports) | 26 | 2,084 | 6 s |
| C3 sentiment expressions | 6 (focus, distraction, stress, relaxation, depression, excitement) | 23 | 133 | 2.5 min |
| C4 gestures at antenna distances | the 21 gestures of C1, RFID only, the antenna at 1.5 m or 3 m, in front or at the side | 17 | 10,620 | 1.5 s |

| Environment | Modalities | Participants | Trials (C1 / C2 / C3) |
|---|---|---|---|
| 1 laboratory | radar, LoRa, RFID, motion capture, IMU (C3); C4 RFID only | 44 | 17,030 (4,193 / 2,084 / 133 / 10,620) |
| 2 living room | radar, IMU; ECG and EEG in the reference subset | 24 | 2,206 (1,492 / 641 / 73) |
| 3 industrial site | radar, IMU | 9 | 1,182 (1,009 / 160 / 13) |

The same participant number is the same person in every environment: U01 to
U44 took part in the laboratory (29 of them in C1 to C3, 17 in C4), U45 to
U61 in the living room and the industrial site, U62 to U68 in the
physiological reference subset (living room).

In C4 the participant repeated the gestures of C1 while the RFID antenna
stood at one of four positions; the position is the node of the zip and its
folder (13: 1.5 m in front, as in C1 to C3, folder `rfid/`; 15: 1.5 m at the
side, `rfid_15/`; 16: 3 m in front, `rfid_16/`; 17: 3 m at the side,
`rfid_17/`; see `meta/nodes.csv`). C4 adds the tags A1, A5 and A9; their places
on the body were not recorded. C4 has no radar, thus its time stamps come from
the RFID reader clock (12 min 35 s ahead of the radar clock).

### Physiological reference subset

Seven new participants (U62 to U68) completed the campaign 3 affective
protocol in the living room while ECG, EEG, IMU, and radar were recorded at
the same time. ECG was recorded with a Shimmer3 ExG unit at 512 Hz with four
disposable electrodes (RA and LA below the right and left clavicle, RL and LL
on the right and left lower abdomen); the files hold two status columns and
the four calibrated channels LA-RA, LL-LA, LL-RA and Vx-RL in mV. EEG was
recorded with a NeuroSky MindWave Mobile headset: one channel at 512 Hz, the
active electrode on the left frontal forehead, reference and ground on the
left ear; raw samples in ADC units, plus the device's signal-quality index,
eight band powers (delta, theta, low and high alpha, low and high beta, low
and mid gamma) and the eSense attention and meditation scores (0-100) at
about 1 Hz. The streams were started and stopped at approximately
coordinated times; the time stamps of the recording computers align them, and
the reference time of a phase is the radar start.

The subset holds the 34 phases in which all four modalities were recorded.
Eight phases were discarded because one modality is missing: the ECG of U64
(E03, E04), U66 (E05, E06) and U67 (E05, E06) had no usable time
information, the radar of U64 E06 was not recorded, and the IMU of U67 E01
was not recorded (`meta/excluded_physio_sessions.csv`). Only 9 of the 13
radars ran in this subset (radars 4, 6, 8 and 12 are absent in most phases),
and some IMU files are empty; the trial table says so for every phase.

Files: `ecg/x_ecg_2_<user>_<E..>_1_<time>.zip` (`ecg.csv`) and
`eeg/x_eeg_2_<user>_<E..>_1_<time>.zip` (`raw_eeg_512hz.csv`,
`summary_1hz.csv`), next to the radar and IMU zips of the same phases. The
rows of `meta/trials_LivingRoom.csv` with `subset = physiological_reference`
carry the ECG and EEG facts; `meta/self_assessment_physio.csv` holds the
per-phase self-ratings (0 to 5) of the seven participants. The physiological
recordings are a per-subject reference for the induced conditions, not a
population-level sample.

| Modality | Sensor | Rate | Content of one trial |
|---|---|---|---|
| `radar` | 13 × TI IWR1443 mmWave radars, 77–81 GHz | about 30 frames/s (ground), 5–7 frames/s (ceiling) | Point clouds: time, x, y, z, signal strength, for each radar |
| `lora` | Semtech SX1276 LoRa node + USRP receiver, 865.5 MHz | 200 Hz features | Amplitude of the baseband signal, its difference, and its variance (20 Hz) |
| `rfid` | 6 Alien AZ 9662 tags on the arms (C4: 9), Impinj Speedway R420 reader | about 50 reads/s in total | One row per tag read: time stamp, tag EPC, RSSI, phase |
| `mocap` | 24 infrared cameras (motion capture) | 100 Hz | Position and rotation of rigid bodies: chest, arms; in C2 and C3 also hips and legs |
| `imu` | 3 Movesense sensors (chest, right arm, left arm; in C2 chest, right arm, right leg) | 104 Hz | Acceleration, angular rate, magnetic field; in the laboratory C3 only |
| `ecg` | Shimmer C9EB | 500 Hz | Two status values and four calibrated differential channels |
| `eeg` | one raw EEG channel | nominal 512 Hz raw; 1 Hz summary | Raw samples, signal quality, attention, meditation, and band features |

In total: 42,758 zips, 1.7 GB; the laboratory alone has 16 million
radar points, 3.9 million RFID reads and about 11 hours of radar recordings.

Segments, with one segment per trial and modality, and one segment per radar
file (each of the 13 radars):

| Campaign (laboratory) | Total segments |
|---|---|
| C1 gestures | 66,167 |
| C2 activities | 31,943 |
| C3 sentiment expressions | 2,258 |
| C4 gestures at antenna distances | 10,620 |
| Total laboratory | 110,988 |
| Living room | 26,947 |
| Industrial site | 13,880 |
| Total | 151,815 |

## 2. Files and names

One zip for each trial and modality, in flat folders (a folder on the Hub
holds at most 10,000 files, thus the C4 antenna positions 15, 16, 17 have
their own folders):

```
radar/   r_radar_1_1_M10_1_20250710163420.zip    radar_00.npz ... radar_12.npz
lora/    5_lora_1_1_M10_1_20250710163420.zip     abs_200Hz.csv, diff_200Hz.csv, var_20Hz.csv
rfid/    13_rfid_1_1_M10_1_20250710163420.zip    rfid.csv   (node 13: the antenna in front at 1.5 m; C1 to C4)
rfid_15/ 15_rfid_1_2_M01_1_20240726143130.zip    rfid.csv   (C4: antenna at the side, 1.5 m; rfid_16/ 3 m front, rfid_17/ 3 m side)
mocap/   14_mocap_1_1_M10_1_20250710163420.zip   mocap.csv
imu/     x_imu_1_3_E01_1_20250720180733.zip      Chest_acc_data.csv, Chest_gyro_data.csv, ...
ecg/     x_ecg_2_62_E01_1_20261002152433.zip      ecg.csv
eeg/     x_eeg_2_62_E01_1_20261002152433.zip      raw_eeg_512hz.csv, summary_1hz.csv
meta/    trials_Lab.csv, trials_LivingRoom.csv, trials_Industry.csv, classes.csv, nodes.csv, packing_log_<Env>.csv
scripts/ readers, loader, download and unpack scripts, demo notebook
```

The file name is `<node>_<modality>_<environment>_<user>_<class>_<repetition>_<time>.zip`:

| Field | Meaning |
|---|---|
| node | Position of the sensor: `r` all radars (the radar number is in the file name inside the zip), `5` LoRa (next to radar 5), `13` RFID antenna, `14` infrared cameras, `x` body-worn IMUs. Positions in `meta/nodes.csv`. |
| environment | `1` laboratory, `2` living room, `3` industrial site |
| user | Participant number |
| class | `M01`–`M21` gestures (C1 and C4), `A01`–`A10` activities, `E01`–`E06` sentiment expressions; names in `meta/classes.csv`. The trial table says which campaign a trial belongs to. |
| repetition | Repetition number of the class by this participant |
| time | Start of the trial, `YYYYMMDDhhmmss`, local time (Europe/Helsinki). The same value in all zips of one trial. |

The available zips of one trial share the same name apart from the node and the modality.

### Radar files

`radar_<n>.npz` holds `points`, an array with one row per detected point (time in
Unix seconds, x, y, z in metres in the frame of the radar, signal strength), and
`frame`, the frame number of each row. The radars remove static reflections. The
transformation to the global frame (origin at the standing point, z up, x
towards radar 5) is in the paper (Section 3) and in `scripts/Radar/read_vis.py`
(`to_global`; views `points` and `animation`), `run_ground.py`, and `run_ceiling.py`.

### Trial table

`meta/trials_<Environment>.csv` has one row per trial with: participant, class, class
name, repetition, the reference time, the height of the ceiling radars (5 m or
3 m), and for each modality whether it exists, its own start time, its duration,
and extra facts (missing radars, number of radar points, RFID reads, rigid
bodies, IMU sensors, empty IMU files).
Physiological rows also include the self-rating, ECG and EEG statistics,
missing internal streams, synchronization basis, and quality notes.
`meta/self_assessment_physio.csv` stores only anonymous participant identifiers.

## 3. Download

The dataset is gated: accept the terms on this page, then log in once
(`hf auth login`). Download a selection with the script in `scripts/download/`:

```bash
python download_rfbehavior.py --campaign C1 --user 1 3 4 --modality radar lora --out ./RF-Behavior
python download_rfbehavior.py --config selection.json --out ./RF-Behavior
python download_rfbehavior.py --campaign C3 --user 62 63 --modality radar imu ecg eeg --out ./RF-Behavior
```

or with the Hub library:

```python
from huggingface_hub import snapshot_download
snapshot_download("Si-Z/RF-Behavior", repo_type="dataset", local_dir="RF-Behavior",
                  allow_patterns=["meta/*", "radar/r_radar_1_1_M*", "lora/5_lora_1_1_M*"])
```

## 4. Load

`scripts/loader/rfbehavior_loader.py` reads the downloaded folder. The calls are
the same as in the OctoNet loader:

```python
from rfbehavior_loader import get_dataset, get_dataloader

config = {"environment": [1], "campaign": ["C1"], "user_list": None, "class_list": None,
          "node_id": [5, 6, 7], "modality": ["radar", "lora", "rfid", "mocap"], "require_all": True}
dataset = get_dataset(config, "RF-Behavior")
sample = dataset[0]
sample["label"], sample["class_name"]            # 9, 'arms swing'
sample["modality_data"]["radar"][5]["points"]    # (N, 5) array
loader = get_dataloader(dataset, batch_size=4, shuffle=True)
```

For the reference subset, request `environment: [2]`, users 62 to 68, and
modalities `radar`, `imu`, `ecg`, and `eeg`. The ECG reader returns status and
channel arrays. The EEG reader returns `raw` and `summary` dictionaries.

`scripts/<Modality>/read_vis.py` reads one trial and draws it: radar point
clouds in the room (3-D animation), LoRa features, RFID tag motion on a body
map, motion-capture skeletons, IMU signals, and class matrices. They need
only numpy, pandas, and matplotlib. They work on the recording layout; one
command puts a download into it:

```bash
python scripts/download/unpack_release.py --download ./RF-Behavior --out ./RF-Behavior_unpacked
python scripts/Radar/read_vis.py --root ./RF-Behavior_unpacked/Lab/Radar --view animation --campaign C2 --user U01 --cls A01
```

`scripts/demo/demo.ipynb` is a notebook in the style of the OctoNet demo:
download with the OctoNet demo's `streaming.py` and its `config.json` (the
copy in `scripts/demo/` differs in three lines, marked), `load_recording`,
`iter_segments`, and `show_keyframe` for every modality
(`scripts/demo/rfb_demo.py`).

Nothing here depends on OctoNet code; only the file-name convention, the
loader calls, and the demo calls follow it, so that the two datasets can be
used side by side.

## 5. License and citation

The data: CC BY-NC-SA 4.0, academic research only. The scripts in `scripts/`:
MIT license (see `scripts/LICENSE`). Please cite:

```bibtex
@article{zuo2025rfbehavior,
  title   = {RF-Behavior: A Multimodal Radio-Frequency Dataset for Activity and Affective Behavior Analysis},
  author  = {Zuo, Si and Song, Yuqing and Han, Jin and Golipoor, Sahar and Liu, Ying and Ma, Xujun and Holme, Petter and Sigg, Stephan},
  journal = {arXiv preprint arXiv:2511.06020},
  year    = {2025}
}
```

Contact: Si Zuo, Aalto University.