DRIFT (Drone-derived Intelligent for Traffic analysis) is an open-source dataset designed to support advanced traffic behavior research using high-resolution drone imagery. It enables accurate vehicle detection, trajectory tracking, and traffic flow analysis across complex urban intersections.
- Provide a large-scale, annotated drone dataset optimized for traffic analysis
- Support urban mobility research with pre-trained models and analytical tools
- Enable multi-scale traffic analysis (microscopic, mesoscopic, and macroscopic)
- To detect and track vehicle instances in high resolution using polygon-based oriented bounding boxes (OBB) at an altitude of 250 meters
- To stabilize video data and real-world orthophoto-mapped trajectories
- To provide 81,699 annotated vehicle trajectories collected across 2.6 km of urban roadways
- To offer object detection/tracking model for customization and the built-in tools for lane-change analysis, time-to-collision (TTC), congestion detection, flow-density analysis, and more
- Site coverage: 9 interconnected urban intersections in Daejeon, South Korea
- Target site: From 99 Daehak-ro to 291 Daehak-ro Daejeon in South Korea
- Target site: From 99 Daehak-ro to 291 Daehak-ro Daejeon in South Korea
- Imagery: 4K drone footage with frame-level annotations
- Trajectory format: Real-world coordinates with speed, acceleration, and heading
- Model: YOLOv11m + ByteTrack with polygon-based OBB detection
| Column | Description |
|---|---|
track_id |
Unique identifier assigned to each vehicle throughout its trajectory |
frame |
Frame index in the video sequence (30 fps) |
center_x, center_y |
Horizontal and vertical positions of the vehicle center, respectively |
width |
Width of the detected vehicle |
height |
Height of the detected vehicle |
angle |
Orientation angle of the vehicle in radians |
x1, y1 |
Coordinates of the front-left corner of the vehicle |
x2, y2 |
Coordinates of the front-right corner of the vehicle |
x3, y3 |
Coordinates of the rear-right corner of the vehicle |
x4, y4 |
Coordinates of the rear-left corner of the vehicle |
confidence |
Confidence score of the detection result (range: 0 to 1) |
class_id |
Object class label (0: bus, 1: car, 2: truck) |
site |
Identifier of the observation site |
lane |
Lane index where the vehicle is currently located |
preceding_id |
Identifier of the vehicle directly ahead |
following_id |
Identifier of the vehicle directly behind |
Note: Position- and size-related values (e.g., coordinates, width, height) are expressed in pixels.
- Frame stabilization
stabilization.mp4
od_small.mp4
https://huggingface.co/datasets/Hj-Lee/The-DRIFT
# DRIFT dataset load
from datasets import load_dataset
dataset = load_dataset("Hj-Lee/The-DRIFT")
# Clone the repository
git clone https://github.com/AIxMobility/The-DRIFT
# Create conda env
conda create -n DRIFT python=3.11 -y
conda activate DRIFT
# Install dependencies
cd The-DRIFT
pip install -r requirements.txt
# Stabilize drone video
sh preprocessing/stabilization.sh
python preprocessing/stabilization.py
# Preprocess the dataset
sh preprocessing/extraction.sh
python preprocessing/extraction.py
# Train detection model
python model/train.py│
├── data/ # Raw and processed drone data
│ ├── csv/ # Frame-level trajectory metadata
│ ├── sample_video/ # Sample drone videos
│ ├── site_images/ # Reference frames in each site
│
├── preprocessing/ # Data extraction and stabilization
│ ├── detect_and_track.py
│ ├── json_to_csv.py
│ ├── lane.py
│ ├── RoI.json
│ ├── extraction.sh
│ ├── extraction.py
│ ├── stabilo.py # Stabilization scripts (Ack.: https://github.com/rfonod/stabilo)
│ ├── default.yaml
│ ├── stabilze_video.py
│ ├── stabilo_utils.py
│ ├── script_utils.py
│ ├── stabilization.sh
│ ├── stabilization.py
│ ├── geoalign_roi.json
│ ├── geoalign_transformation.ipynb
|
├── model/ # Annotation data and model training
│ ├── test/
│ ├── train/
│ ├── valid/
│ ├── data.yaml
│ ├── drone_data.yaml
│ ├── train.py
│ ├── best.pt
|
├── utils/ # Utility scripts for data handling
│ ├── convert.py
│ ├── video_to_frame.py
│
├── vis/ # Visualization scripts and tools
│
├── notebooks/
│ ├── data_exploration.ipynb
│ ├── performance_analysis.ipynb
│
├── requirements.txt
├── README.md
└── LICENSE
| Scale | Description | Image |
|---|---|---|
| Microscopic | Lane Change (LC) |
|
| Microscopic | Time-to-Collision (TTC) |
|
| Mesoscopic | Flow-Density Diagram |
|
| Mesoscopic | Time-Space Diagram |
|
| Macroscopic | Speed Heatmap |
|
This project is based on the Stabilo repository by Robert Fonod, licensed under the MIT License.
Certain parts have been adapted and modified to better suit the needs of this project.
GitHub: https://github.com/rfonod/stabilo
@software{fonod2025stabilo,
author = {Fonod, Robert},
license = {MIT},
month = apr,
title = {Stabilo: A Comprehensive Python Library for Video and Trajectory Stabilization with User-Defined Masks},
url = {https://github.com/rfonod/stabilo},
doi = {10.5281/zenodo.12117092},
version = {1.0.1},
year = {2025}
}
DOI: 10.5281/zenodo.12117092
@misc{fonod2025advanced,
title={Advanced computer vision for extracting georeferenced vehicle trajectories from drone imagery},
author={Robert Fonod and Haechan Cho and Hwasoo Yeo and Nikolas Geroliminis},
year={2025},
eprint={2411.02136},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2411.02136},
doi={https://doi.org/10.48550/arXiv.2411.02136}
}
If you use this project in your academic research, commercial products, or any published material, please acknowledge its use by citing it.
@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},
}