---
license: mit
language:
- en
tags:
- computer-vision
- pose-estimation
- drowsiness-detection
- driver-monitoring
pretty_name: FatigueSense Training Data
---

# FatigueSense Training Data

Domain-specific recordings and derived training artifacts for the FatigueSense
fatigue-detection pipeline. This dataset supports **YOLO11n-pose retraining**
(upper-body keypoints) and **BiGRU temporal model** training (1 Hz feature windows).

## Contents

| Path | Used by | Description |
|------|---------|-------------|
| `pose/` | `model_architecture.train_yolo_pose` | YOLO-pose images + labels (5 kpts: nose, ears, shoulders). Pseudo-labeled with `yolo11n-pose.pt`. |
| `pose/dataset.yaml` | Ultralytics training | Portable config (`path: .`); 70/15/15 train/val/test **by video** (no subject leakage). |
| `temporal/features/` | `model_architecture.train_temporal_model` | Stage 2 Parquets — one file per video, 17-dim per-second features. |
| `temporal/raw_probs/` | Reproducibility | Stage 1 per-frame eye/mouth probs + pose keypoints. |
| `videos/` | Optional | 25 raw source clips (upload with `--include-videos`). |
| `manifest.json` | — | Split stems, feature schema, file counts (auto-generated on upload). |

## Provenance

1. **Videos** — local `videos/` directory (25 clips, driver/desk POV).
2. **Pose** — `scripts/pose/extract_frames.py` → `scripts/pose/pseudo_label.py` → `data/pose/`.
3. **Temporal** — `scripts/temporal/extract_probs.py` → `scripts/temporal/aggregate_features.py` → `data/temporal/`.

Temporal labels are **bootstrap heuristics** over the same features (PERCLOS, yawn rate, blink duration, posture), not human focus scores. Replace when annotated labels exist.

## Feature schema (17 dimensions)

`perclos`, `blink_rate_bpm`, `mean_blink_duration`, `eye_closure_variance`,
`yawn_rate_per_min`, `mean_yawn_duration`, `mouth_open_ratio`, `mean_p_eye`,
`mean_p_mouth`, `head_pitch`, `head_roll`, `shoulder_tilt`, `head_size_ratio`,
`head_motion_energy`, `head_drift_y`, `posture_drift`, `kpt_visibility`.

See `fatigue_pipeline/constants.py` for definitions.

## Download

```python
from huggingface_hub import snapshot_download

root = snapshot_download("Jlords32/FatigueSense", repo_type="dataset")
# pose:     {root}/pose
# temporal: {root}/temporal/features
```

## Citation

If you use this data, cite the FatigueSense project repository.