COSMOS Trajectory Dataset

Description

This dataset comprises real-world pedestrian and vehicle trajectories collected from the COSMOS Testbed. The trajectories are captured at an intersection in New York City using a high-elevation camera using the tracking-by-detection paradigm with a fine-tuned YOLOv8 object detection model and the BoT-SORT tracking algorithm.

Complete trajectories of 176,000 pedestrians and 215,000 vehicles, suitable for statistical analyses. In the context of trajectory forecasting, the dataset offers 510k pedestrian and 250k vehicle trajectories collected over a 30-day period. They are processed into 10k segments of 20-frame (8 seconds) scenes for model training and evaluation.

This dataset is designed to support research in trajectory forecasting, traffic simulation, and urban mobility analysis.

Organization

The dataset is structured as follows:

.
├── split                                           # train-test splits
│   └── cosmos_l20_f2p5                             #   └─ cosmos_l[SEQ_LEN]_f[FPS]
│       ├── test                                    #       ├─ test set
│       ├── train                                   #       ├─ train set
│       ├── val                                     #       └─ valid set
│       ├── cosmos_231212_l20_f2p5_train.npy
│       ├── cosmos_231212_l20_f2p5_val.npy
│       └── cosmos_240109_l20_f2p5_test.npy
└── README.md

Each split contains a number of .npy files, each representing a scene of length SEQ_LEN, each file is named as cosmos_[DATE]_l[SEQ_LEN]_f[FPS]_[HORIZON], where each field are

DATE: the date of collection
SEQ_LEN: maximum length of the sequence
FPS: FPS of the data
HORIZON: prediction horizon (index of the first frame of the predictions)

Each .npy file is a 2D-array with 5 columns, i.e. each line of the array contains:

frame_id: index of the current frame
track_id: unique tracking ID of the object
label: class label of the object (0 - vehicle, 1 - pedestrian)
pos_x: x-coordinate of the bouding box center
pos_y: y-coordinate of the bouding box center

For example, the content of split/cosmos_l20_f2p5/train/cosmos_231212_l20_f2p5_51054441684.npy reads

	frame_id	track_id	label	pos_x	pos_y
0	5.105444e+10	5692.0	1.0	617.261447	251.848858
1	5.105444e+10	5692.0	1.0	616.453705	253.075640
2	5.105444e+10	5692.0	1.0	616.374630	252.099720
3	5.105444e+10	5692.0	1.0	615.510800	252.186663
4	5.105444e+10	5692.0	1.0	614.769500	253.377940
...	...	...	...	...	...

357 rows × 5 columns

Citation

If you use this dataset, please cite the following paper:

C. Zang, M. K. Turkcan, G. Zussman, J. Ghaderi, and Z. Kostic, “Data-Driven Traffic Simulation for an Intersection in a Metropolis,” presented at the CVPR Workshop POETS, Seattle, WA, USA, Jun. 2024. Available: https://arxiv.org/abs/2408.00943.

@misc{zang2024data,
  author       = {Zang, Chengbo and Turkcan, Mehmet Kerem and Zussman, Gil and Ghaderi, Javad and Kostic, Zoran},
  title        = {Data-Driven Traffic Simulation for an Intersection in a Metropolis},
  howpublished = {Presented at the CVPR Workshop POETS},
  month        = jun,
  year         = {2024},
  address      = {Seattle, WA, USA},
  note         = {Available at \url{https://arxiv.org/abs/2408.00943}},
  url          = {https://arxiv.org/abs/2408.00943}
}

Acknowledgements

This work was supported in part by NSF grants CNS1827923 and EEC-2133516, NSF grant CNS-2038984 and corresponding support from the Federal Highway Administration (FHA), NSF grant CNS-2148128 and by funds from federal agency and industry partners as specified in the Resilient & Intelligent NextG Systems (RINGS) program, and ARO grant W911NF2210031.