AutomatumData/automatum-data-highway-area-pools

AUTOMATUM Highway Area Pools v1.0.0

Longer trajectories by flying two drones side by side and fusing their tracks into one scene.

Each scene in this dataset was recorded by two drones at the same time, hovering over neighbouring sections of a German Autobahn. Each video is processed on its own by the AUTOMATUM pipeline: image stabilisation, CNN detection, georeferencing, tracking, perspective correction and vehicle dynamics. The two recordings are then pooled: we estimate the time offset between them, associate the vehicles seen by both drones in the overlap area and merge them into one consistent trajectory set.

On average, pooling makes the observed road section 1.9× longer than the best single drone. Vehicles that are associated across both drones are followed for more than a kilometre. This applies to 86 % of the vehicles at A9 Offenbau and to 34–39 % at the other two scenes. With a single drone, almost no vehicle is observed over 1 km.

scene	road	tracks (pooled)	road length, best single drone	road length, pooled	median track length, single	median track length, pooled	tracks > 1 km, single	tracks > 1 km, pooled	vehicle-km	estimated time offset
A6_SchwabachWest	A6, junction Schwabach-West	1103	607 m	1170 m	615 m	642 m	0.0 %	38.8 %	894	-0.58 s
A9_OffenbauRastanlage	A9, service area Offenbau	1127	664 m	1278 m	675 m	1298 m	0.1 %	86.4 %	1339	+0.12 s
A9_LentingSolarpark	A9, Lenting (solar park section)	1721	640 m	1238 m	636 m	640 m	0.0 %	34.5 %	1386	+0.20 s

Scene notes In all three scenes the two drone footprints overlap by only about 25–30 m along the road.

• A9 Offenbau (service area): 94 % of the vehicles in the common time span are associated across the two drones (998 of 1065).
• A6 Schwabach-West (junction with ramps): 55 % are associated (503 of 915). The others remain separate tracks per drone.
• A9 Lenting. The recording is from November 2024 and shows a construction-site traffic layout: traffic in both directions runs on one carriageway ("4+0"). This is a rare scenario type. 54 % of the vehicles are associated (614 of 1146).

We are still investigating why the association rate differs so much between the scenes, given that the overlap is similar. A wider overlap is planned for future flights.

"Road length" is the 0.5–99.5 % extent of all vehicle positions along the main road axis. "Track length" is the distance driven per vehicle while it is observed.

What is in the dataset

data/<scene>/pooled/tracks.parquet         one row per vehicle and frame (29.97 Hz), fused scene
data/<scene>/pooled/tracks_meta.parquet    one row per vehicle (class, length, width, start/end time)
data/<scene>/pooled/scene.json             reference point (WGS84 + UTM), drone footprints (`frame_borders_utm`), object counts, versions
data/<scene>/single/drone_1|drone_2/...    the same tables for each drone on its own (for comparison)
data/<scene>/pooled/automatum/             native AUTOMATUM format: dynamicWorld.json + staticWorld.xodr/.rd5
data/<scene>/single/drone_*/automatum/     the same per drone (works with `pip install openautomatumdronedata`)
data/<scene>/images/drone_*_center_image.jpg  stabilised reference frame of each drone
data/<scene>/sync/time_sync_curve.json     time-offset search curve used for the pool (deviation + #matched vehicles per offset)
data/A9_LentingSolarpark/sync/time_sync_curve_wide.json   the ±20 s landscape for Lenting (final pool: ±2 s fine search)
metrics/pool_metrics.json                  all numbers of the table above, plus histograms
manifest.json                              file list with sha256

Signals (tracks.parquet)

column	unit	meaning
scene_id, track_id	–	scene and vehicle UUID
t_s, frame	s, –	time since scene start, frame index at 29.97 fps
x_vec, y_vec	m	position relative to the scene's UTM reference point (scene.json)
x_utm_m, y_utm_m	m	absolute position in WGS84 / UTM zone 32N (EPSG:32632)
z_m (pooled)	m	road-surface height above sea level (DHHN2016) at the vehicle position, from the official DGM1 terrain model and airborne LiDAR incl. bridge decks
psi_vec	rad	heading in the UTM frame, wrapped to (−π, π] in the Parquet tables (unwrapped in dynamicWorld.json)
vx_vec, vy_vec	m/s	velocity along / across the vehicle's heading (vehicle frame)
ax_vec, ay_vec	m/s²	acceleration along / across the vehicle's heading
jerk_x_vec, jerk_y_vec, curvature_vec	m/s³, 1/m	derived dynamics

Object classes: car, van, truck, carWithTrailer, bus, motorcycle.

# native AUTOMATUM tooling
from openautomatumdronedata.dataset import droneDataset   # pip install openautomatumdronedata setuptools
ds = droneDataset("data/A9_OffenbauRastanlage/pooled/automatum")
print(len(ds.dynWorld.get_list_of_dynamic_objects()))

from datasets import load_dataset
pooled = load_dataset("AutomatumData/highway-area-pools", "pooled", split="train")
# or directly with pandas
import pandas as pd
df = pd.read_parquet("hf://datasets/AutomatumData/highway-area-pools/data/A9_OffenbauRastanlage/pooled/tracks.parquet")

How the drones are synchronised

The two videos are not frame-synchronised. The pooling step searches for the time offset Δt at which the vehicles inside the overlap area of both frames match best. For every candidate offset it records the mean position deviation and the number of vehicles that could be associated. sync/time_sync_curve.json contains the whole curve. In all three scenes the correct offset is a sharp, unambiguous peak: 330–420 associated vehicles within about ±0.2 s, compared with typically 20–60 elsewhere.

Two lessons from this release:

• Use the matched-vehicle count, not deviation alone. Minimising the deviation alone can lock onto spurious offsets that are supported by only a handful of vehicles. An earlier processing of A6 Schwabach-West had selected +11.8 s; the correct offset is −0.58 s.
• Do not trust metadata start times. The recording start times disagree with the true offset by 70 s in one scene, A9 Lenting, while the true offset is +0.19 s. A wide search with a coarse-to-fine step is the safer approach.

Processing versions and limitations (please read)

• Pipeline versions. The single-drone tracks come from the AUTOMATUM pipeline with image stabilisation 3.3 and detection 2.7. In this release the downstream steps were recomputed where the road model had changed. These are not the latest internal production versions (3.4 / 3.0). Each scene.json lists the exact CalculationStepVersions.
• Height. The x/y positions come from the production georeferencing, a plane fitted through ground control points. The pooled tracks also carry z_m, the height of the road surface at each sample. We took it from the Bavarian DGM1 terrain model and airborne LiDAR, with bridge decks extracted from the point cloud. The heights from the two drones agree to within a few centimetres. A fully terrain-aware re-projection of x/y improved the cross-drone consistency only at the scene with the strongest relief, so x/y are left unchanged in this version. Geodata: Bayerische Vermessungsverwaltung – www.geodaten.bayern.de, CC BY 4.0, data modified.
• Maps. staticWorld.xodr/.rd5 are the OpenDRIVE/RoadRunner models of each drone location. The pooled scene is expressed in the frame of drone 1, so its map covers the drone-1 section. Pool-wide maps are in preparation. A9 Offenbau has no map yet; one will follow in a later version.
• Licence change. Earlier AUTOMATUM releases on Hugging Face use CC BY-ND 4.0. This release uses CC BY-NC 4.0, so that researchers may modify the data, for example re-segment or fuse it.
• Ends of the frame. Tracks start and end at the outer border of the pooled frame. Vehicles entering or leaving at junctions and service areas are cut where they leave the observed area.
• Hand-over between drones. At the sample where a merged track passes from one drone to the other, the time step can deviate from 1/29.97 s by a few hundredths of a second (sub-frame time offset). If you derive speeds from positions, use t_s rather than the frame index, or use the provided vx_vec.
• Lenting map. The OpenDRIVE/RoadRunner model for A9 Lenting shows the road layout before the construction works. It does not represent the 4+0 layout of the recording.
• No raw video. No raw video and no personal data are included. Vehicles are represented only by their class, bounding-box size and motion.

Geodata attribution

The height values (z_m), the maps shown on this page and our 3D renderings use official open geodata: Datenquelle: Bayerische Vermessungsverwaltung – www.geodaten.bayern.de, Daten verändert (CC BY 4.0, https://creativecommons.org/licenses/by/4.0/). Products used: DOP20, DGM1, airborne laser point clouds, LoD2 buildings.

License

CC BY-NC 4.0. Free for non-commercial research and teaching, with attribution. For commercial use (OEM/Tier-1 validation, simulation, scenario databases), for custom recordings at your own locations, or for larger multi-drone pools, contact Automatum Data GmbH: info@automatum-data.com · https://www.automatum-data.com

Citation

A paper describing the pooling method is in preparation. Until it is out, please cite:

@misc{automatum_highway_area_pools_1_0_0,
  title  = {AUTOMATUM Highway Area Pools: multi-drone fused highway trajectories},
  author = {Spannaus, Paul and Zechel, Peter and Lenz, Kilian},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/AutomatumData/highway-area-pools}},
  note   = {Version 1.0.0}
}