AutomatumData/automatum-data-full-highway

---
language:
- en
- de
license: cc-by-nd-4.0
tags:
- autonomous-driving
- traffic-analysis
- trajectory-prediction
- drone-data
- automatum
- open-drive
- json
- highway
- ALKS
- benchmark
- openscenario
pretty_name: "Automatum Data: Full Highway Drone Dataset"
task_categories:
- time-series-forecasting
- object-detection
size_categories:
- 100K<n<1M
---


# Automatum Data: Full Highway Drone Dataset

[](https://automatum-data.com)
[](https://openautomatumdronedata.readthedocs.io)
[](https://pypi.org/project/openautomatumdronedata/)
[](https://creativecommons.org/licenses/by-nd/4.0/)
[](doc/IV21_Automatumd_Full_Drone_Dataset.pdf)

## Introduction

The **Automatum Data Full Highway Dataset** is a large-scale collection of high-precision vehicle trajectory data extracted from **30 hours of drone video** capturing **12 characteristic highway scenes** along the German A9 Autobahn. With approximately **200,000 tracked vehicles** and over **80,000 km of cumulative trajectory data**, this dataset represents one of the most comprehensive open drone-based highway datasets available.

The processing pipeline incorporates deep learning (Faster R-CNN) for detection and LOESS filtering for stabilization, achieving an exceptionally low **relative velocity error of less than 0.2%**, validated against instrumented reference vehicles.


## Dataset at a Glance

| Metric | Value |
|--------|-------|
| **Scenario Type** | Highway (straight segments) |
| **Recordings** | 114 |
| **Locations** | 11 along the A9 Autobahn |
| **Total Duration** | ~30 hours |
| **Total Vehicles Tracked** | ~200,000 |
| **Total Distance** | ~80,000 km |
| **Velocity Error** | < 0.2% (validated with reference vehicles) |
| **Coordinate System** | UTM Zone 32U |
| **FPS** | 29.97 |
| **License** | CC BY-ND 4.0 |


## Repository Structure

```
automatum-data-full-highway/
├── README.md # This file
├── doc/ # Documentation images, logo, paper
├── example_scripts/ # Ready-to-use Python analysis scripts
├── Sample_Data/ # One recording unpacked for quick preview
│ └── hw-a9-appershofen-001-.../
│ ├── dynamicWorld.json
│ ├── staticWorld.xodr
│ ├── recording.html
│ └── img/
└── automatum_data_full_highway_drone_dataset.zip # All 114 recordings as archive (~4 GB)
```

> **Quick Preview:** Browse `Sample_Data/` to explore the data structure before downloading the full archive (~4 GB). The sample recording can be loaded directly with the `openautomatumdronedata` Python library.

## KPI Comparison with Established Datasets

| Metric | **Automatum Data** | highD Dataset | NGSIM (US-101 / I-80) |
|--------|-------------------|---------------|------------------------|
| **Total Duration** | **30 hours** | 16.5 hours | ~1.5 hours |
| **Total Vehicles** | **~200,000** | 110,000 | ~thousands |
| **Total Distance** | **~80,000 km** | 45,000 km | limited segments |
| **Source / Perspective** | Drone / Aerial | Drone / Aerial | Fixed Cameras & Drones |
| **Error / Accuracy** | **< 0.2% velocity** | typically < 10 cm | Known clipping issues |
| **Static Description** | **OpenDRIVE XODR** | simple XML/CSV | Basic annotations |
| **Data Format** | **JSON** | CSV | CSV |
| **Object Relationships** | **Built-in (TTC, TTH)** | Must compute | Must compute |
| **OpenSCENARIO** | **Available on request** | No | No |


## Recording Locations

The 114 recordings span 11 locations along the German A9 Autobahn:

| Location | Recordings | Description |
|----------|-----------|-------------|
| Denkendorf | 36 | Major section with high traffic density |
| Stammham | 16 | Mixed traffic scenarios |
| Appershofen | 14 | Varied speed profiles |
| Dunzendorf | 11 | Characteristic highway flow |
| Kinding | 9 | Multi-lane segments |
| Brunn | 9 | Standard highway traffic |
| Hausen | 7 | Diverse driving patterns |
| Untermässing | 6 | Rural highway section |
| Heppberg Park | 3 | Near rest area |
| Apperszell | 2 | Additional coverage |
| Ingolstadt Nord | 1 | Urban highway approach |

## Data Structure

Each recording folder follows the naming convention `hw-a9-{location}-{sequence}-{uuid}` and contains:

```
hw-a9-appershofen-001-uuid/
├── dynamicWorld.json # Trajectories, velocities, accelerations, bounding boxes
├── staticWorld.xodr # Road geometry in OpenDRIVE format
├── recording_name.html # Interactive metadata overview (Bokeh)
└── img/ # (may contain visualizations)
```

### dynamicWorld.json

The core data file contains for each tracked vehicle:

- **Position vectors**: `x_vec`, `y_vec` — UTM coordinates over time
- **Velocity vectors**: `vx_vec`, `vy_vec` — in m/s
- **Acceleration vectors**: `ax_vec`, `ay_vec` — in m/s²
- **Jerk vectors**: `jerk_x_vec`, `jerk_y_vec`
- **Heading**: `psi_vec` — orientation angle
- **Lane assignment**: `lane_id_vec`, `road_id_vec` — linked to XODR
- **Object dimensions**: `length`, `width`
- **Object relationships**: `object_relation_dict_list` — front/behind/left/right neighbors
- **Safety metrics**: `ttc_dict_vec` (Time-to-Collision), `tth_dict_vec` (Time-to-Headway)
- **Lane distances**: `distance_left_lane_marking`, `distance_right_lane_marking`


### staticWorld.xodr

OpenDRIVE 1.6 format file defining:

- Road network topology and geometry
- Lane definitions with widths and types
- Speed limits (typically 100 km/h, unlimited sections)
- Road markings and surface properties


### Key Metrics Explained


## Quick Start

### Installation

```bash
pip install openautomatumdronedata
```

### Load and Explore

```python
from openautomatumdronedata.dataset import droneDataset
import os

# Point to one recording folder
path = os.path.abspath("hw-a9-appershofen-001-uuid")
dataset = droneDataset(path)

# Access dynamic world
dynWorld = dataset.dynWorld

print(f"UUID: {dynWorld.UUID}")
print(f"Duration: {dynWorld.maxTime:.1f} seconds")
print(f"Frames: {dynWorld.frame_count}")
print(f"Vehicles: {len(dynWorld)}")

# Get all vehicles visible at t=5.0s
objects = dynWorld.get_list_of_dynamic_objects_for_specific_time(5.0)
for obj in objects[:5]:
speed_kmh = ((obj.vx_vec[0]**2 + obj.vy_vec[0]**2)**0.5) * 3.6
print(f" {obj.UUID} ({obj.type}) — {speed_kmh:.1f} km/h")
```

### Using with Hugging Face

```python
from huggingface_hub import snapshot_download, hf_hub_download
import zipfile, os

# Option 1: Download only the sample for a quick look (~200 MB)
local_path = snapshot_download(
repo_id="AutomatumData/automatum-data-full-highway",
repo_type="dataset",
allow_patterns=["Sample_Data/**"]
)

# Option 2: Download the full archive (~4 GB)
archive = hf_hub_download(
repo_id="AutomatumData/automatum-data-full-highway",
filename="automatum_data_full_highway_drone_dataset.zip",
repo_type="dataset"
)
# Extract
with zipfile.ZipFile(archive, 'r') as z:
z.extractall("automatum_data_full_highway")

# Load with openautomatumdronedata
from openautomatumdronedata.dataset import droneDataset
dataset = droneDataset("automatum_data_full_highway/hw-a9-appershofen-001-d8087340-8287-46b6-9612-869b09e68448")
print(f"Vehicles: {len(dataset.dynWorld)}")
```

### Batch Processing All Recordings

```python
from openautomatumdronedata.dataset import droneDataset
import os
import json

base_path = "path/to/automatum_data_full_highway_drone_dataset"

stats = []
for folder in sorted(os.listdir(base_path)):
full_path = os.path.join(base_path, folder)
if not os.path.isdir(full_path) or not folder.startswith("hw-"):
continue

dataset = droneDataset(full_path)
dw = dataset.dynWorld

stats.append({
"recording": folder,
"vehicles": len(dw),
"duration_s": dw.maxTime,
"frames": dw.frame_count,
})
print(f"{folder}: {len(dw)} vehicles, {dw.maxTime:.0f}s")

# Save summary
with open("dataset_summary.json", "w") as f:
json.dump(stats, f, indent=2)
```

## Example Scripts

See the `example_scripts/` folder for ready-to-use analysis scripts:

- **`01_lane_changes.py`** — Analyze lane change behavior across all vehicles
- **`02_heatmap_density.py`** — Generate traffic density heatmaps
- **`03_high_acceleration.py`** — Detect high-acceleration events

## Research Paper

The methodology and validation of this dataset are described in our peer-reviewed publication:

> **AUTOMATUM DATA: Drone-based highway dataset for development and validation of automated driving software for research and commercial applications**
> Paul Spannaus, Peter Zechel, Kilian Lenz
> *IEEE Intelligent Vehicles Symposium (IV), 2021*

The paper is included in this repository: [`doc/IV21_Automatumd_Full_Drone_Dataset.pdf`](doc/IV21_Automatumd_Full_Drone_Dataset.pdf)

Key findings from the paper:

- Processing pipeline validated with instrumented reference vehicles
- Relative velocity error < 0.2%
- Deep learning detection (Faster R-CNN) combined with LOESS filtering
- High-precision UTM world coordinate mapping
- Standardized OpenDRIVE export for seamless integration with simulation tools

## Research Use & Extended Data Pool

**These publicly available datasets are intended exclusively for research purposes.**

This dataset, while comprehensive, is still an excerpt from the full **Automatum Data Pool** containing over **1,000 hours of processed drone video** across highways, intersections, roundabouts, and urban scenarios. For commercial use or access to further datasets, including OpenSCENARIO exports, please contact us via our website:

**[automatum-data.com](https://automatum-data.com)**

## Citation

If you use this dataset in your research, please cite:

```bibtex
@inproceedings{spannaus2021automatum,
title={AUTOMATUM DATA: Drone-based highway dataset for development and validation of automated driving software},
author={Spannaus, Paul and Zechel, Peter and Lenz, Kilian},
booktitle={IEEE Intelligent Vehicles Symposium (IV)},
year={2021}
}
```

## License

This dataset is licensed under [Creative Commons Attribution-NoDerivatives 4.0 International (CC BY-ND 4.0)](https://creativecommons.org/licenses/by-nd/4.0/).

## Contact

- **Website**: [automatum-data.com](https://automatum-data.com)
- **Email**: [e-mail]
- **HuggingFace**: [AutomatumData](https://huggingface.co/AutomatumData)
- **Documentation**: [openautomatumdronedata.readthedocs.io](https://openautomatumdronedata.readthedocs.io)

---
语言：
- en
- de
许可协议：知识共享署名-禁止演绎4.0国际许可协议（CC BY-ND 4.0）
标签：
- 自动驾驶（autonomous-driving）
- 交通分析（traffic-analysis）
- 轨迹预测（trajectory-prediction）
- 无人机数据（drone-data）
- Automatum
- OpenDRIVE（open-drive）
- JSON
- 高速公路（highway）
- 自动车道保持系统（ALKS）
- 基准测试（benchmark）
- OpenSCENARIO
数据集名称："Automatum Data：全高速公路无人机数据集"
任务类别：
- 时间序列预测（time-series-forecasting）
- 目标检测（object-detection）
数据规模：100,000 < n < 1,000,000
---


# Automatum Data：全高速公路无人机数据集

[](https://automatum-data.com)
[](https://openautomatumdronedata.readthedocs.io)
[](https://pypi.org/project/openautomatumdronedata/)
[](https://creativecommons.org/licenses/by-nd/4.0/)
[](doc/IV21_Automatumd_Full_Drone_Dataset.pdf)

## 简介

**Automatum Data全高速公路数据集**是一套大规模高精度车辆轨迹数据集，源自对德国A9高速公路沿线12个典型高速公路场景的30小时无人机航拍视频提取得到。该数据集共追踪了约**20万辆车辆**，累计轨迹数据总长度超过**8万公里**，是目前公开的基于无人机采集的最全面的高速公路数据集之一。

该数据集的处理管线集成了用于目标检测的深度学习模型（Faster R-CNN）与用于数据稳定化的LOESS滤波，经配备测试仪器的参考车辆验证，其相对速度误差仅低于**0.2%**。


## 数据集概览

| 指标 | 数值 |
|--------|-------|
| **场景类型** | 高速公路（直线路段） |
| **录制片段数** | 114 |
| **采集点位** | 沿A9高速公路的11个点位 |
| **总时长** | 约30小时 |
| **总追踪车辆数** | 约20万辆 |
| **总轨迹里程** | 约8万公里 |
| **速度误差** | <0.2%（经参考车辆验证） |
| **坐标系** | 通用横轴墨卡托坐标系32U带（UTM Zone 32U） |
| **帧率** | 29.97 |
| **许可协议** | CC BY-ND 4.0 |


## 仓库结构

automatum-data-full-highway/
├── README.md # 本说明文件
├── doc/ # 文档用图片、Logo与论文
├── example_scripts/ # 可直接使用的Python分析脚本
├── Sample_Data/ # 单条录制片段解压预览包
│ └── hw-a9-appershofen-001-.../
│ ├── dynamicWorld.json
│ ├── staticWorld.xodr
│ ├── recording.html
│ └── img/
└── automatum_data_full_highway_drone_dataset.zip # 全部114条录制片段压缩包（约4 GB）

> **快速预览**：下载完整压缩包（约4 GB）前，可先浏览`Sample_Data/`文件夹了解数据结构。该示例录制片段可直接通过`openautomatumdronedata` Python库加载。

## 与主流数据集的KPI对比

| 指标 | **Automatum Data** | highD数据集 | NGSIM（US-101 / I-80） |
|--------|-------------------|---------------|------------------------|
| **总时长** | **30小时** | 16.5小时 | 约1.5小时 |
| **总车辆数** | **~20万辆** | 11万辆 | 数千辆 |
| **总轨迹里程** | **~8万公里** | 4.5万公里 | 仅有限路段 |
| **数据来源/采集视角** | 无人机/航拍视角 | 无人机/航拍视角 | 固定摄像头与无人机 |
| **误差/精度** | **速度误差<0.2%** | 通常<10cm | 存在已知的裁剪问题 |
| **静态场景描述** | **OpenDRIVE XODR格式** | 简单XML/CSV格式 | 基础标注 |
| **数据格式** | **JSON** | CSV | CSV |
| **目标关系** | **内置（碰撞时间TTC、车头时距TTH）** | 需自行计算 | 需自行计算 |
| **OpenSCENARIO** | **可按需获取** | 不支持 | 不支持 |


## 录制点位

114条录制片段覆盖沿德国A9高速公路的11个点位：

| 点位 | 录制片段数 | 描述 |
|----------|-----------|-------------|
| 登肯多夫（Denkendorf） | 36 | 高交通密度核心路段 |
| 施塔姆哈姆（Stammham） | 16 | 混合交通场景 |
| 阿珀斯霍芬（Appershofen） |14 | 多样速度分布 |
| 敦岑多夫（Dunzendorf） |11 | 典型高速公路车流 |
| 金丁（Kinding） |9 | 多车道路段 |
| 布伦（Brunn） |9 | 标准高速公路交通 |
| 豪森（Hausen） |7 | 多样化驾驶行为 |
| 翁特梅辛（Untermässing） |6 | 乡村高速公路路段 |
| 赫普贝格公园（Heppberg Park） |3 | 临近休息区 |
| 阿珀斯策尔（Apperszell） |2 | 补充覆盖路段 |
| 因戈尔施塔特北（Ingolstadt Nord） |1 | 城市高速入口路段 |

## 数据结构

每条录制片段的文件夹命名遵循`hw-a9-{location}-{sequence}-{uuid}`格式，包含以下文件：

hw-a9-appershofen-001-uuid/
├── dynamicWorld.json # 轨迹、速度、加速度与边界框信息
├── staticWorld.xodr # OpenDRIVE格式的道路几何信息
├── recording_name.html # 交互式元数据概览（基于Bokeh可视化库）
└── img/ # （可选包含可视化结果）

### dynamicWorld.json

核心数据文件包含每台追踪车辆的以下信息：
- **位置向量**：`x_vec`、`y_vec` — 随时间变化的通用横轴墨卡托坐标系（UTM）坐标
- **速度向量**：`vx_vec`、`vy_vec` — 单位：米每秒（m/s）
- **加速度向量**：`ax_vec`、`ay_vec` — 单位：米每二次方秒（m/s²）
- **加加速度向量**：`jerk_x_vec`、`jerk_y_vec`
- **航向角**：`psi_vec` — 车辆朝向角度
- **车道分配**：`lane_id_vec`、`road_id_vec` — 与XODR文件关联
- **目标尺寸**：`length`、`width`
- **目标关系**：`object_relation_dict_list` — 前后左右相邻车辆
- **安全指标**：`ttc_dict_vec`（碰撞时间，Time-to-Collision, TTC）、`tth_dict_vec`（车头时距，Time-to-Headway, TTH）
- **车道距离**：`distance_left_lane_marking`、`distance_right_lane_marking` — 与左右车道线的距离


### staticWorld.xodr

采用OpenDRIVE 1.6格式的文件，包含以下信息：
- 道路网络拓扑与几何结构
- 车道定义，包含车道宽度与类型
- 限速（通常为100 km/h，部分路段无上限）
- 道路标线与路面属性


### 关键指标说明


## 快速入门

### 安装

bash
pip install openautomatumdronedata

### 加载与探索数据

python
from openautomatumdronedata.dataset import droneDataset
import os

# 指向单条录制片段文件夹
path = os.path.abspath("hw-a9-appershofen-001-uuid")
dataset = droneDataset(path)

# 获取动态世界数据
dynWorld = dataset.dynWorld

print(f"UUID: {dynWorld.UUID}")
print(f"时长: {dynWorld.maxTime:.1f} 秒")
print(f"总帧数: {dynWorld.frame_count}")
print(f"车辆总数: {len(dynWorld)}")

# 获取t=5.0s时刻可见的所有车辆
objects = dynWorld.get_list_of_dynamic_objects_for_specific_time(5.0)
for obj in objects[:5]:
speed_kmh = ((obj.vx_vec[0]**2 + obj.vy_vec[0]**2)**0.5) * 3.6
print(f" {obj.UUID} ({obj.type}) — {speed_kmh:.1f} km/h")

### 使用Hugging Face（拥抱脸）加载数据

python
from huggingface_hub import snapshot_download, hf_hub_download
import zipfile, os

# 选项1：仅下载示例数据快速预览（约200 MB）
local_path = snapshot_download(
repo_id="AutomatumData/automatum-data-full-highway",
repo_type="dataset",
allow_patterns=["Sample_Data/**"]
)

# 选项2：下载完整压缩包（约4 GB）
archive = hf_hub_download(
repo_id="AutomatumData/automatum-data-full-highway",
filename="automatum_data_full_highway_drone_dataset.zip",
repo_type="dataset"
)
# 解压文件
with zipfile.ZipFile(archive, 'r') as z:
z.extractall("automatum_data_full_highway")

# 使用openautomatumdronedata加载数据
from openautomatumdronedata.dataset import droneDataset
dataset = droneDataset("automatum_data_full_highway/hw-a9-appershofen-001-d8087340-8287-46b6-9612-869b09e68448")
print(f"车辆总数: {len(dataset.dynWorld)}")

### 批量处理所有录制片段

python
from openautomatumdronedata.dataset import droneDataset
import os
import json

base_path = "path/to/automatum_data_full_highway_drone_dataset"

stats = []
for folder in sorted(os.listdir(base_path)):
full_path = os.path.join(base_path, folder)
if not os.path.isdir(full_path) or not folder.startswith("hw-"):
continue

dataset = droneDataset(full_path)
dw = dataset.dynWorld

stats.append({
"recording": folder,
"vehicles": len(dw),
"duration_s": dw.maxTime,
"frames": dw.frame_count,
})
print(f"{folder}: {len(dw)} 辆车辆, {dw.maxTime:.0f}秒")

# 保存数据集摘要
with open("dataset_summary.json", "w") as f:
json.dump(stats, f, indent=2)

## 示例脚本

可查看`example_scripts/`文件夹中的现成分析脚本：
- **`01_lane_changes.py`** — 分析所有车辆的变道行为
- **`02_heatmap_density.py`** — 生成交通密度热力图
- **`03_high_acceleration.py`** — 检测高加速度事件

## 研究论文

该数据集的处理方法与验证细节已发表在同行评审期刊论文中：

> **AUTOMATUM DATA：面向自动驾驶软件开发与验证的无人机采集高速公路数据集**
> Paul Spannaus, Peter Zechel, Kilian Lenz
> *IEEE智能车辆研讨会（IV 2021）*

论文已包含在本仓库中：[`doc/IV21_Automatumd_Full_Drone_Dataset.pdf`](doc/IV21_Automatumd_Full_Drone_Dataset.pdf)

论文核心结论：
- 处理管线经配备测试仪器的参考车辆验证
- 相对速度误差<0.2%
- 结合深度学习检测（Faster R-CNN）与LOESS滤波
- 高精度通用横轴墨卡托坐标系（UTM）世界坐标映射
- 标准化OpenDRIVE导出格式，可无缝集成至仿真工具

## 研究用途与扩展数据集池

**本公开数据集仅用于学术研究用途。**
尽管该数据集已较为全面，但它仅为完整**Automatum Data数据集池**的一部分，该数据集池包含超过1000小时的无人机航拍视频，覆盖高速公路、交叉口、环岛与城市场景。如需商业使用或获取更多数据集（含OpenSCENARIO导出文件），请通过官网联系我们：
**[automatum-data.com](https://automatum-data.com)**

## 引用

若您在研究中使用本数据集，请引用以下文献：

bibtex
@inproceedings{spannaus2021automatum,
title={"AUTOMATUM DATA: Drone-based highway dataset for development and validation of automated driving software"},
author={Spannaus, Paul and Zechel, Peter and Lenz, Kilian},
booktitle={IEEE Intelligent Vehicles Symposium (IV)},
year={2021}
}

## 许可协议

本数据集采用[知识共享署名-禁止演绎4.0国际许可协议（CC BY-ND 4.0）](https://creativecommons.org/licenses/by-nd/4.0/)进行许可。

## 联系方式

- **官网**：[automatum-data.com](https://automatum-data.com)
- **邮箱**：[e-mail]
- **Hugging Face（拥抱脸）**：[AutomatumData](https://huggingface.co/AutomatumData)
- **文档**：[openautomatumdronedata.readthedocs.io](https://openautomatumdronedata.readthedocs.io)