---
license: cc-by-4.0
pretty_name: MoCap2Radar v3
size_categories:
  - n<1K
tags:
  - radar
  - micro-doppler
  - motion-capture
  - time-series
---

# MoCap2Radar v3

Synchronised optical motion-capture (53 markers, 240 Hz) and 5.8 GHz radar I/Q
(256 Hz) recordings of a walking subject. Used in *What Physics do Data-Driven MoCap-to-Radar Models Learn?* (Kevin Chen, Kenneth W. Parker, Anish Arora; 2026 IEEE Radar Conference, RadarConf26; [arXiv:2605.00018](https://arxiv.org/abs/2605.00018)).

## Recordings

| recording | split | duration (s) | windows (256 / hop 32) |
|---|---|---|---|
| DiagonalLong1 | train | 118.6 | 942 |
| DiagonalLong1Fast | train | 60.4 | 476 |
| DiagonalLong2 | train | 103.7 | 822 |
| DiagonalLong2Fast | train | 54.9 | 432 |
| DiagonalLong3 | train | 117.1 | 930 |
| DiagonalLong3Fast | train | 71.3 | 563 |
| DiagonalLong4 | val | 119.8 | 951 |
| DiagonalLong4Fast | val | 81.1 | 641 |
| RandomWalk1 | eval | 303.0 | 2416 |
| RandomWalk2 | eval | 301.4 | 2404 |

## Files

- `raw/v3/mocap/<recording>.csv` — 53 markers × (x, y, z) in mm, world frame of the
  capture volume, 240 Hz. Vicon export: first row is the subject label
  (`Subject1_9_20`, anonymised), second row the column headers
  (`Subject1_9_20\<MARKER><T-X|Y|Z>`). Rows after the `Radar9_20` sentinel are a
  second Vicon subject, the four markers on the radar board, and are dropped by the loader.
- `raw/v3/radar/<recording>.csv` — first column is a beacon column (dropped);
  remaining columns are I/Q samples at 256 Hz.
- `config.yaml` — STFT parameters (nperseg 256, noverlap 224), 50 ms startup delay,
  mocap sample rate, and the split definition above.
- `scalers.npz` — `StandardScaler` mean / scale fitted on the train split only:
  `mocap` [159], `mocap_local` [159], `mocap_doppler` [53], `radar_sxx` [1],
  `radar_iq_centered` [2]. Stored as plain arrays (`<name>_mean`, `<name>_scale`).
- `SHA256SUMS` — checksums of every file above.

## Preprocessing

Regenerate the model-ready arrays with `python src/preproc.py` from the code
repository (<CODE URL>). The pipeline: load radar I/Q and mocap, apply the startup
delay, trim both streams to a common duration, upsample mocap to 256 Hz, convert
mocap to radar-relative coordinates (radar centroid at the origin), window with
nperseg 256 / hop 32, compute the two-sided complex STFT of the radar I/Q and take
magnitude in dB, fit the scalers on the train split and apply them to all splits.
`scalers.npz` lets you skip the fitting step and check your regeneration against it.

## Consent and ethics

<TO FILL: subject consent statement; approval reference if any>

## Citation

```bibtex
@inproceedings{chen2026whatphysics,
  title     = {What Physics do Data-Driven MoCap-to-Radar Models Learn?},
  author    = {Chen, Kevin and Parker, Kenneth W. and Arora, Anish},
  booktitle = {2026 IEEE Radar Conference (RadarConf26)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2605.00018}
}
```