AutomatumData/automatum-data-full-highway

Automatum Data: Full Highway Drone Dataset

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

pip install openautomatumdronedata

Load and Explore

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

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

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

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

Citation

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

@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).

Contact

• Website: automatum-data.com
• Email: info@automatum-data.com
• HuggingFace: AutomatumData
• Documentation: openautomatumdronedata.readthedocs.io