rfonod/songdo-traffic

Songdo Traffic: High Accuracy Georeferenced Vehicle Trajectories from a Large-Scale Study in a Smart City

Songdo Traffic is a large-scale dataset of ~700,000 georeferenced vehicle trajectories extracted from high-altitude bird's-eye view (BEV) drone footage. The trajectories carry WGS84 coordinates, kinematics (speed, acceleration), estimated vehicle dimensions, class, and lane/road assignment, sampled at 29.97 points per second. They were captured during a large-scale multi-drone urban traffic monitoring experiment over the Songdo International Business District, South Korea, and constitute one of the most extensive aerial traffic datasets publicly available. Songdo Traffic is the trajectory dataset produced by the Geo-trax pipeline and the associated publication.

📦 The data is hosted on Zenodo, not on Hugging Face. This page is a documentation mirror and index card. The full dataset (trajectories, orthophotos, segmentations, master frames, and sample videos; ~70 GB) lives immutably on Zenodo under concept DOI 10.5281/zenodo.13828384 (always resolves to the latest version; the current release is v2, record 17924857). See Access the dataset below for how to download it.

Dataset at a glance

Property	Value
Modality	Georeferenced vehicle trajectories extracted from 4K (3840 × 2160) RGB BEV video
Trajectories	~700,000 unique trajectories
Temporal resolution	29.97 points/second (per the 29.97 FPS source video)
Classes	4: 0 car (incl. vans), 1 bus, 2 truck, 3 motorcycle
Coordinate systems	Orthophoto pixels · local Cartesian (KGD2002 / Central Belt 2010, EPSG:5186) · WGS84 (EPSG:4326)
Per-vehicle attributes	Position, speed, acceleration, estimated length/width, class, lane, road section, visibility
Acquisition	Fleet of 10 drones over 20 intersections (A–U), 29.97 FPS, 140–150 m altitude, DJI Mavic 3
Collection dates	October 4–7, 2022
Location	Songdo International Business District, South Korea
Total size	~70 GB (trajectories ~41 GB + orthophotos, master frames, segmentations, sample videos)
License	CC BY 4.0

Trajectory CSV schema

Each trajectory CSV contains one row per vehicle observation with the following fields:

Column	Format	Type	Description
Vehicle_ID	1, 2, …	int	Unique identifier within the CSV
Local_Time	hh:mm:ss.sss	str	Korean local time (GMT+9), ISO 8601
Drone_ID	1–10	int	Source drone identifier
Ortho_X, Ortho_Y	pixels (1 d.p.)	float	Orthophoto image coordinates
Local_X, Local_Y	meters (2 d.p.)	float	KGD2002 / Central Belt 2010 (EPSG:5186)
Latitude, Longitude	DD (7 d.p.)	float	WGS84 geographic (EPSG:4326)
Vehicle_Length, Vehicle_Width	meters (2 d.p.)	float	Estimated physical dimensions
Vehicle_Class	0–3	int	0 = car/van, 1 = bus, 2 = truck, 3 = motorcycle
Vehicle_Speed	km/h (1 d.p.)	float	Computed with Gaussian smoothing
Vehicle_Acceleration	m/s² (2 d.p.)	float	Derived from smoothed speed
Road_Section	N_G	str	Node and lane-group identifier
Lane_Number	1, 2, …	int	Lane position (1 = leftmost)
Visibility	0/1	bool	Fully (1) / partially (0) visible in frame

Access the dataset

The dataset is published and versioned on Zenodo. Please download it from the canonical source; it is not redistributed on Hugging Face:

➡️ zenodo.org/records/17924857  ·  DOI 10.5281/zenodo.13828384 (always resolves to the latest version; the current release is v2, record 17924857)

Files on Zenodo:

File(s)	Size	Contents
YYYY-MM-DD_<ID>.zip (80 archives)	16.2–360.2 MB each	Trajectory CSVs per intersection (A–U) and day; 10 sessions each (AM1–AM5, PM1–PM5)
orthophotos.zip	1.8 GB	8000 × 8000 px orthophoto cut-outs per intersection + georeferencing parameters
master_frames.zip	248.6 MB	Reference frames + homography parameters for georeferencing reproducibility
segmentations.zip	24.9 KB	Road/lane polygon definitions (CSV) per intersection
sample_videos.zip	26.8 GB	29 sample 4K UHD drone clips (first 60 s of the PM5 session, Oct 7, 2022)
README.md	small	Dataset documentation
LICENSE.txt	small	CC BY 4.0 terms

v2 update: version 2 adds the master frames and georeferencing parameters (for full reproducibility of the coordinate transforms) and updated coordinate-transformation documentation.

Download with zenodo_get:

pip install zenodo_get
zenodo_get 10.5281/zenodo.13828384      # fetches all files for the latest version

Or directly with requests:

import requests

# Resolve the record's files via the Zenodo API, then download each one.
record = requests.get("https://zenodo.org/api/records/17924857").json()
for f in record["files"]:
    url = f["links"]["self"]
    print("Downloading", f["key"], f"({f['size'] / 1e9:.2f} GB)")
    with requests.get(url, stream=True) as r, open(f["key"], "wb") as out:
        for chunk in r.iter_content(chunk_size=1 << 20):
            out.write(chunk)

Relationship to Geo-trax

Songdo Traffic is the trajectory dataset produced by Geo-trax, a pipeline that extracts georeferenced vehicle trajectories from high-altitude drone footage. The trajectories are generated end-to-end by detecting and tracking vehicles, stabilizing the footage, georeferencing against orthophotos, and estimating per-vehicle kinematics, dimensions, and lane assignments. The default YOLOv8s detector driving the pipeline is published separately as the Hugging Face model rfonod/geo-trax, trained and validated on the companion detection dataset rfonod/songdo-vision.

Note: These trajectories are extracted automatically and carry estimation artifacts (fragmentation, kinematic/dimension estimation errors, occasional lane misassignment). See Known limitations below and the model card before using them.

Known limitations

The trajectories are extracted automatically and carry estimation artifacts. In brief:

• Trajectory fragmentation: motorcycles fragment in complex infrastructure; some trucks are underrepresented; drone technical issues occasionally split recordings mid-session.
• Vehicle ID ambiguities: the largest Vehicle_ID in a CSV does not equal the total count of unique vehicles.
• Kinematics: speed/acceleration are affected by detection inaccuracies, stabilization artifacts, interpolation, and smoothing.
• Dimension estimation: unreliable for stationary or non-axially-aligned vehicles.
• Lane/section assignment: perspective effects can misassign tall vehicles (trucks/buses) to adjacent lanes.

See the Zenodo record and the publication for the full discussion of limitations.

Related datasets and resources

• Songdo Vision: companion vehicle-detection (annotated image) dataset: 10.5281/zenodo.13828407 · HF rfonod/songdo-vision
• Geo-trax detector: HF model rfonod/geo-trax
• Source video recordings (not open access): 10.5075/EPFL.20.500.14299/253923
• Publication: Transportation Research Part C (2025): 10.1016/j.trc.2025.105205 · arXiv:2411.02136
• Software: Geo-trax: github.com/rfonod/geo-trax · demo video

Citation

If you use this dataset, please cite the associated publication:

@article{fonod2025advanced,
  title   = {Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery},
  author  = {Fonod, Robert and Cho, Haechan and Yeo, Hwasoo and Geroliminis, Nikolas},
  journal = {Transportation Research Part C: Emerging Technologies},
  volume  = {178},
  pages   = {105205},
  year    = {2025},
  doi     = {10.1016/j.trc.2025.105205}
}

Please also cite the dataset itself via its Zenodo record:

@dataset{fonod2025songdotraffic,
  author    = {Fonod, Robert and Cho, Haechan and Yeo, Hwasoo and Geroliminis, Nikolas},
  title     = {Songdo Traffic: High Accuracy Georeferenced Vehicle Trajectories from a Large-Scale Study in a Smart City},
  year      = {2025},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.13828384},
  url       = {https://doi.org/10.5281/zenodo.13828384}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license; see the LICENSE file for the full terms. You are free to share and adapt the material for any purpose, provided you give appropriate credit.