README.md

---
pretty_name: DriveDNA-Sample
license: other
license_name: drivedna-research-license
license_link: https://huggingface.co/datasets/HenryYHW/DriveDNA
language:
  - en
task_categories:
  - time-series-forecasting
  - other
tags:
  - naturalistic-driving
  - driving-style
  - driver-identification
  - autonomous-driving
  - CAN-bus
  - multimodal
  - time-series
size_categories:
  - n<1K
arxiv: "2607.23822"
---

# 🧬 DriveDNA-Sample

A small, directly browsable slice of [**DriveDNA**](https://huggingface.co/datasets/HenryYHW/DriveDNA).
Every drive is one folder holding one concatenated forward-view **mp4** and one decoded
**csv** on the same clock — no manifests to join, no segment files to stitch.

**63 drives · 56.2 h · 6.55 GB video · 0.55 GB signals · 2,048,504 rows · 14 car models**

Driver pseudonyms (`driver_078`, `drive_004`, …) are **identical to those in the full
DriveDNA release**, so anything you find here can be traced straight back to the main
dataset's splits, embeddings, and index tables.

---

## Why these drivers

The sample is not a random subset. It is built around the two comparisons that DriveDNA's
benchmark is designed to separate — *the same person in different cars* versus
*different people in the same car*.

### 🚗 One driver, many vehicles — `driver_078`

`driver_078` logged **14 different car models**, from a Toyota Camry to a Tesla Model X to a
VW Tiguan. Their behavioural signature has to be read *through* fourteen different vehicle
platforms, each with its own steering geometry, powertrain response, and CAN signal set.
This is the axis that isolates driver identity from vehicle dynamics.

| Car model | Drives | Hours |
|---|---|---|
| `HONDA_ACCORD_HYBRID_2018` | 4 | 3.18 |
| `HONDA_CIVIC` | 1 | 0.60 |
| `HYUNDAI_IONIQ_5` | 4 | 4.96 |
| `HYUNDAI_IONIQ_5_2022` | 4 | 3.56 |
| `KIA_EV6` | 4 | 1.31 |
| `KIA_NIRO_EV_2ND_GEN` | 4 | 1.38 |
| `KiaNiro2023` | 4 | 6.59 |
| `TESLA_AP3_MODEL_3` | 4 | 3.08 |
| `TESLA_MODEL_X` | 4 | 3.11 |
| `TOYOTA_CAMRY_2021` | 4 | 4.48 |
| `TOYOTA_CAMRY_TSS2` | 4 | 4.39 |
| `TOYOTA_RAV4_2023` | 4 | 1.83 |
| `TOYOTA_RAV4_TSS2_2023` | 4 | 4.24 |
| `VOLKSWAGEN_TIGUAN_MK2` | 4 | 3.88 |
| **Total** | **53** | **46.59** |

### 👥 Many drivers, one vehicle — the shared `HONDA_CIVIC`

Four different people drove **the same Honda Civic**. Vehicle platform is held fixed, so
whatever separates these drives is the person behind the wheel, not the car.

| Driver | Drives | Hours | CSV schema |
|---|---|---|---|
| `driver_174` | 5 | 6.90 | `drivedna_v1` |
| `driver_369` | 1 | 1.00 | `drivedna_v1` |
| `driver_372` | 1 | 1.05 | `drivedna_v1` |
| `driver_427` | 3 | 0.70 | `legacy_gnss_removed` |
| **Total** | **10** | **9.65** | |

`driver_078` also appears once on the Civic (`HONDA_CIVIC/driver_078`), so the two groups
overlap on a single shared platform.

---

## 📁 Layout

```
Dataset/<CAR_MODEL>/<driver_XXX>/<drive_YYY>/
                                 ├── drive_YYY.mp4    # concatenated forward video
                                 └── drive_YYY.csv    # decoded signals, same clock
index.csv                                             # one row per drive
```

```
Dataset/TESLA_MODEL_X/driver_078/drive_021/drive_021.mp4
Dataset/TESLA_MODEL_X/driver_078/drive_021/drive_021.csv
Dataset/HONDA_CIVIC/driver_372/drive_001/drive_001.mp4
Dataset/HONDA_CIVIC/driver_372/drive_001/drive_001.csv
```

`drive_YYY` numbering is **per driver**, exactly as in the full release: `driver_078`'s
drives run `drive_001…drive_240` across all fourteen of their car models.

---

## 🎬 Video

Each `.mp4` is the drive's ~60-second `qcamera` segments concatenated in order.

| | |
|---|---|
| Codec | H.264, **stream-copied** — not re-encoded, bit-identical to source |
| Resolution | 526 × 330 |
| Frame rate | 20 fps |
| Audio | **removed** (`-an`) |
| Container | MP4, `+faststart` |

```bash
ffmpeg -f concat -safe 0 -fflags +genpts -i segments.txt \
       -c copy -an -movflags +faststart drive_YYY.mp4
```

In 12 of the 63 files the **final frame** is truncated and decoders emit one
`error while decoding MB …` warning at the very end. This comes from the source
recordings — the logger stopped mid-frame — and is present in the original segments
as well; stream copying preserves it rather than introducing it. Everything before
the last frame decodes cleanly.

---

## 📊 Signals

### `drivedna_v1` — 27 columns @ 10 Hz (60 of 63 drives)

Identical to `raw_signals_csv/<driver_XXX>/<drive_YYY>.csv` in the full DriveDNA release.

| Group | Columns |
|---|---|
| Time | `time_s` |
| Ego motion | `vEgo`, `aEgo`, `vEgoCluster`, `yawRate`, `yaw_rate` |
| Steering | `steeringAngleDeg`, `steeringRateDeg`, `steeringPressed`, `actual_curvature`, `curv_measured`, `slip` |
| Pedals | `gas`, `gasPressed`, `brake`, `brakePressed` |
| Lead vehicle | `leadOne_status`, `leadOne_dRel`, `leadOne_vLead`, `leadOne_vRel` |
| Lane | `laneLeft_y`, `laneRight_y` |
| Indicators | `leftBlinker`, `rightBlinker` |
| Automation | `cruiseState_enabled` (OEM ACC), `cs_enabled` (openpilot) |
| Human mask | `is_human` |

**`is_human` is the column that matters.** ADAS-engaged frames are kept in place rather
than cut out, so the timeline stays continuous against the video. Filter to
`is_human == 1` for human-controlled driving:

```python
import pandas as pd
d = pd.read_csv("Dataset/TESLA_MODEL_X/driver_078/drive_021/drive_021.csv")
human = d[d.is_human == 1]          # video time = d.time_s, unchanged
```

### `legacy_gnss_removed` — 3 columns @ 20 Hz (`driver_427`, 3 drives)

`driver_427`'s Civic drives predate the current pipeline and no raw logs survive for them,
only a pre-rendered mp4 and a thin CSV. They are included so the shared-Civic group is
complete, but they carry **`t`, `yaw_ned`, `vEgo` only** — no CAN channels, no `is_human`
mask. The original `lat` / `lon` / `alt` columns were **stripped**, matching DriveDNA's
policy of releasing no GNSS coordinates.

These three drives (`driver_427/drive_003…005`) exist **only in this sample** — the full
DriveDNA release covers `driver_427` through their two Ford Mustang Mach-E drives, which
do have raw logs. Check `csv_schema` in `index.csv` before assuming columns exist.

---

## 🗂️ `index.csv`

One row per drive: `group`, `car_model`, `driver`, `drive`, `n_video_segments`,
`video_s`, `csv_rows`, `csv_cols`, `csv_s`, `csv_hz`, `csv_schema`, `mp4_MB`, `csv_MB`.

```python
ix = pd.read_csv("index.csv")
ix[ix.group == "one_driver_many_vehicles"].car_model.nunique()   # 14
```

---

## ⏱️ Clock alignment

Video and CSV both start at drive time zero, so CSV row `i` sits at `i / csv_hz` seconds
into the mp4. Recording occasionally stops a moment later on one stream than the other:
the median `csv_s − video_s` gap is **0.27 s**, and the largest is **75.9 s** (one trailing
segment logged after the camera stopped). Use `video_s` and `csv_s` in `index.csv` when
exact end-of-drive behaviour matters.

---

## 🔐 Privacy

- Drivers appear only as `driver_XXX`; raw device identifiers are not distributed.
- **No GPS.** No coordinates in any schema; the legacy CSV had its GNSS columns removed.
- **No audio.** Microphone tracks present in some source recordings are dropped.
- Forward-view video only — no cabin or driver-facing camera.
- Original collection files were never modified in producing this release.

---

## 📄 Full dataset & citation

The complete corpus — 465 drivers, 115 car models, 4,121 drives, 975 h, plus frozen video
embeddings, benchmark splits, maneuver annotations, and the evaluation harness — lives at
[**HenryYHW/DriveDNA**](https://huggingface.co/datasets/HenryYHW/DriveDNA).

```bibtex
@article{drivedna2026,
  title  = {DriveDNA: A Large-Scale Multimodal Naturalistic Driving Dataset
            and Benchmark for Driving Style Identification},
  author = {Wang, Yuhang and Li, Lingyao and Kontar, Wissam and Wen, Jason and Zhou, Hao},
  year   = {2026},
  eprint = {2607.23822},
  archivePrefix = {arXiv}
}
```