README.md

metadata

license: cc-by-4.0
task_categories:
  - object-detection
  - other
tags:
  - trajectory
  - autonomous-driving
  - traffic
  - drone
  - urban
  - vehicle-tracking
  - uav
  - intersection
pretty_name: The DRIFT Open Dataset
size_categories:
  - 10M<n<100M
language:
  - en

The DRIFT Open Dataset

DRIFT (Drone-derived Intelligence for Traffic analysis) is a large-scale open dataset of urban vehicle trajectories captured by drone at 250 m altitude across 9 interconnected intersections in Daejeon, South Korea.

Code & Pipeline: github.com/AIxMobility/The-DRIFT
Paper: arXiv:2504.11019

---

Dataset at a Glance

Trajectories	81,699 annotated vehicle trajectories
Coverage	2.6 km of continuous urban roadway (9 intersections)
Location	Daejeon, South Korea (99–291 Daehak-ro)
Altitude	~250 m
Resolution	4K drone footage
Frame rate	30 fps
Vehicle classes	Bus · Car · Truck
Detection model	YOLOv11m + ByteTrack (OBB)

---

Load the Dataset

Note: Due to site-specific columns (transformed_center_x/y in Site F/G only), load_dataset() without streaming=True will raise a schema error. Use one of the methods below.

from datasets import load_dataset

# Option 1: Streaming — quick inspection, no full download
dataset = load_dataset("Hj-Lee/The-DRIFT", split="train", streaming=True)
sample = next(iter(dataset))
print(list(sample.keys()))

# Option 2: Load a single site as pandas (consistent schema within one site)
# Setting HF_TOKEN speeds up downloads but is not required
dataset = load_dataset("Hj-Lee/The-DRIFT", data_files="B/*.csv", split="train")
import pandas as pd
df = dataset.to_pandas()

---

Dataset Columns

All position and size values are in pixel coordinates (4K resolution).
To convert to meters: 1 pixel ≈ 0.065 m at 250 m altitude (site-dependent; see preprocessing/geoalign_transformation.ipynb in the GitHub repo for exact homography).

Column	Type	Description
track_id	float	Unique vehicle identifier per video
frame	int	Frame index (30 fps)
center_x	float	Vehicle center X position (pixels)
center_y	float	Vehicle center Y position (pixels)
width	float	Bounding box width (pixels)
height	float	Bounding box height (pixels)
angle	float	Vehicle orientation (radians)
x1, y1	float	Front-left corner of oriented bounding box (pixels)
x2, y2	float	Front-right corner (pixels)
x3, y3	float	Rear-right corner (pixels)
x4, y4	float	Rear-left corner (pixels)
confidence	float	Detection confidence score (0–1)
class_id	float	Vehicle class: 0=bus · 1=car · 2=truck
site	string	Observation site (e.g., "Site B")
lane	string	Lane code (e.g., "B3")
preceding_id	float	Track ID of the vehicle directly ahead
following_id	float	Track ID of the vehicle directly behind
timestamp	string	Timestamp (ISO 8601)
transformed_center_x	float	Geo-aligned center X (pixels, homography-transformed) — Site F, G only
transformed_center_y	float	Geo-aligned center Y (pixels, homography-transformed) — Site F, G only

Site F & G include transformed_center_x/y columns derived from homography matching defined in preprocessing/geoalign_roi.json. Other sites have these columns as NaN. See the GitHub repo for the transformation pipeline.

---

Example Usage

Filter by site

from datasets import load_dataset
import pandas as pd

# Load a single site (B has 20 drone files, A–I have 13–20 each)
dataset = load_dataset("Hj-Lee/The-DRIFT", data_files="B/*.csv", split="train")
df = dataset.to_pandas()
print(f"Site B: {df['track_id'].nunique()} vehicles, {len(df)} rows")

Visualizations

Microscopic	Lane Change (LC)

Mesoscopic	Macroscopic

---

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

You are free to use, share, and adapt this dataset for any purpose, including commercial use, as long as appropriate credit is given.

Copyright © 2025 Hyejin Lee, AIxMobility Lab, KAIST (Korea Advanced Institute of Science and Technology)
& SAIL (Soonchunhyang Artificial Intelligence Lab), Soonchunhyang University

---

Citation

@misc{lee2025driftopendatasetdronederived,
  title         = {DRIFT open dataset: A drone-derived intelligence for traffic analysis in urban environment},
  author        = {Hyejin Lee and Seokjun Hong and Jeonghoon Song and Haechan Cho and Zhixiong Jin
                   and Byeonghun Kim and Joobin Jin and Jaegyun Im and Byeongjoon Noh and Hwasoo Yeo},
  year          = {2025},
  eprint        = {2504.11019},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2504.11019}
}