README.md

---
dataset_info:
  features:
  - name: event_id
    dtype: int64
  - name: label
    dtype:
      class_label:
        names:
          '0': crash
          '1': near_miss
          '2': normal_driving
  - name: vehiclesize
    dtype: string
  - name: gyro_is_hd
    dtype: bool
  - name: gsensor
    list:
      list: float32
      length: 3
  - name: gyro
    list:
      list: float32
      length: 3
  - name: gps_speed
    list: float32
  splits:
  - name: train
    num_bytes: 5310515525
    num_examples: 137954
  - name: validation
    num_bytes: 1046078370
    num_examples: 27175
  - name: test
    num_bytes: 930673262
    num_examples: 24174
  download_size: 7741526897
  dataset_size: 7287267157
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: cc-by-nc-4.0
pretty_name: VZCrash
size_categories:
- 100K<n<1M
tags:
- timeseries
---


# VZCrash Dataset

VZCrash is a dataset containing telemetry of ego-vehicle crashes. It includes 100 Hz tri-axial accelerometer (in g) and gyroscope (in deg/s) data and 1 Hz GPS-derived speed (in km/h). We provide these signals for almost 190,000 unique events, including above 31,000 verified positive events (crashes).

This dataset complements the paper *VZCrash: A Large-Scale IMU Dataset of Ego-Vehicle Crashes.* Accepted and to be presented in the 2026 IEEE International Conference on Intelligent Transportation Systems (ITSC).

<img src="vzcrash.png" width="800" >

## Dataset overview

### Data
VZCrash is comprised of 100 Hz tri-axial accelerometer (measured in g) and gyroscope (measured in deg/s) data and 1 Hz GPS-derived speed data (measured in km/h).
Each event covers a 16-second window.  
The Inertial Measurement Units (IMUs) that recorded the acceleration and angular velocity data periodically calibrate themselves based on the direction of movement of the vehicles, so that the X-axis points forward,
the Y-axis points leftward and the Z-axis points upward and measures 1g at rest. Note that in some events the axis alignment can be suboptimal.

### Metadata

#### Labels
For each event, we provide a category label with the following possible values:
```
'0': crash
'1': near_miss
'2': normal_driving
```
Each event has been reviewed by three expert human reviewers, trained for this task, with access to dashcam footage in addition to the telemetry data. The final labels are the result of a 3-way consensus.

We define a *crash* as an event involving the ego-vehicle that satisfies at least one of the following criteria:

- **Unintentional contact**: Any physical impact between the ego-vehicle and external entities, such as other vehicles, animals, or infrastructure (e.g., traffic barriers), independently from the fault attribution.
- **Accidental departure**: Unintentional departure from the established roadway or drivable surface.

We define a *near-miss* as any traffic situation requiring a rapid evasive maneuver by the ego-vehicle to successfully avoid a crash. In our paper's experiment, near-misses are used as *negative* examples.

For further details, please refer to the paper.

#### Gyroscope resolution
If you plan to use gyroscope signals, please note that the precision of the angular velocity differ depending on the firmware version of the device.
All gyroscope signals are stored in deg/s, but they differ in resolution:

| Type | `gyro_is_hd` | Resolution | Events |
|------|--------------|------------|--------|
| Standard | `False` | 1 deg/s  | 103,548 |
| High-definition | `True` | 0.001 deg/s  | 85,755 |

The `gyro_is_hd` column in the dataset indicates which resolution applies to each event.


#### Vehicle size
We provide the ego-vehicle size category when known. Events where this information is unavailable are marked as `"Unknown"`.

| `vehiclesize` | Events |
|---------------|--------|
| Light | 27,615 |
| Medium | 67,937 |
| Heavy | 30,967 |
| Unknown | 62,784 |

Our classification can be roughly mapped to the [FHWA Vehicle Classification](https://www.fhwa.dot.gov/policyinformation/tmguide/tmg_2013/vehicle-types.cfm) scheme as follows:
- light-duty vehicles include passenger cars, corresponding to class 2 of the FHWA;
- mid-duty vehicles include vans, pickups and small trucks, corresponding to class 3 and 5 of the FHWA;
- and heavy-duty vehicles include medium and large-size trucks, corresponding to class 6-13 of the FHWA .

## Try it yourself
Clone the repository containing the dataset over HTTPS with: 

```bash
git clone https://huggingface.co/datasets/vzc-research-chapter/VZCrash
```

or over ssh with

```bash
git clone git@hf.co:datasets/vzc-research-chapter/VZCrash
```

You need git-lfs or Xet for this to download the dataset files.
If you want to clone without large files (just their pointers) use `GIT_LFS_SKIP_SMUDGE=1`.

In `explore.ipynb` you can find a demo notebook that shows how to access the data and plot some signals.
In `pr_curve.ipynb` you can find a notebook that lets you reproduce the PR curve for a simple physical baseline
to perform crash detection that we use in the paper.

### License
This dataset is released under the [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) license.

### Citation
If you use this dataset, please cite the accompanying paper:

- **Plain text:** Bianconcini, T., Monteagudo, H. P., Pjetri, A., Trinci, T., & Taccari, L. "VZCrash: A Large-Scale IMU Dataset of Ego-Vehicle Crashes", 2026 IEEE International Conference on Intelligent Transportation Systems (ITSC).
- **BibTeX :** 
```bibtex
@inproceedings{bianconcini2026vzcrash,
  title={VZCrash: A Large-Scale IMU Dataset of Ego-Vehicle Crashes},
  author={Bianconcini, Tommaso and Monteagudo, Henrique Pi{\~n}eiro and Pjetri, Aurel and Trinci, Tomaso and Taccari, Leonardo},
  year={2026}
}
 ```