IDD-PeD: Pedestrian Intention and Trajectory Prediction in Unstructured Traffic Environments

IDD-PeD is a pedestrian behavior dataset for intention and trajectory prediction in unstructured traffic environments. It accompanies the paper "Pedestrian Intention and Trajectory Prediction in Unstructured Traffic Using IDD-PeD" (ICRA 2025).

• Project page: https://cvit.iiit.ac.in/research/projects/cvit-projects/iddped
• Code (baselines + dataset interface): https://github.com/Ruthvik9/IDD-PeD
• Baseline checkpoints: included in this repo under checkpoints/

License note: This card defaults to cc-by-nc-4.0. Update the license field above if your dataset uses a different license.

What's in this repository

This repository hosts the annotations for IDD-PeD. The raw videos are large and are distributed separately from IIIT servers (links below).

annotations/            # Pedestrian spatial + behavioral + scene annotations (CVAT XML, per set/video) + annotated-frame CSVs
annotations_vehicle/    # Vehicle annotations (CVAT XML)
data_cache/
  iddp_database.pkl     # Pre-built database cache used by the dataset interface (iddped_interface_traj.py)
checkpoints/            # Pretrained baseline checkpoints (see "Baseline checkpoints" below)

Baseline checkpoints

Pretrained checkpoints for all baselines are included under checkpoints/:

Folder	Task	Notes
checkpoints/intention/	Pedestrian crossing intention	Image-based; variants: c3d, conv_lstm, i3d, mask_pcpa, pcpa, sf_gru, single_rnn, static
checkpoints/piefull/	PIE trajectory	Sub-models: intention/, speed/, trajectory/
checkpoints/mtn/	MTN trajectory	epoch_latest.pth
checkpoints/bitrap/	BiTraP trajectory	epoch_latest.pth
checkpoints/sgnet/	SGNet trajectory	epoch_latest.pth

Download just the checkpoints with:

hf download Ruthvik9/IDD-PeD --repo-type dataset --include "checkpoints/*" --local-dir .

See the GitHub README for per-model setup and the exact test/train commands.

Downloading the videos

The videos are distributed as 9 tar archives. Download them and arrange them under data/IDDPedestrian/videos/gopro/ as expected by the code:

https://cvit.iiit.ac.in/images/datasets/IDDPed/Videos/gp_set_0001.tar
https://cvit.iiit.ac.in/images/datasets/IDDPed/Videos/gp_set_0002.tar
https://cvit.iiit.ac.in/images/datasets/IDDPed/Videos/gp_set_0003.tar
https://cvit.iiit.ac.in/images/datasets/IDDPed/Videos/gp_set_0004.tar
https://cvit.iiit.ac.in/images/datasets/IDDPed/Videos/gp_set_0005.tar
https://cvit.iiit.ac.in/images/datasets/IDDPed/Videos/gp_set_0006.tar
https://cvit.iiit.ac.in/images/datasets/IDDPed/Videos/gp_set_0007.tar
https://cvit.iiit.ac.in/images/datasets/IDDPed/Videos/gp_set_0008.tar
https://cvit.iiit.ac.in/images/datasets/IDDPed/Videos/gp_set_0009.tar

A convenience script (download_videos.sh) is provided in the GitHub repo.

Train / test split

• Train: gp_set_0001, gp_set_0002, gp_set_0004, gp_set_0006, gp_set_0007 (3284 pedestrians)
• Test: gp_set_0003, gp_set_0005, gp_set_0008, gp_set_0009 (1632 pedestrians)

Roughly a 70/30 split.

Annotation types

The dataset provides five annotation families for pedestrians requiring the ego-vehicle's attention:

1. Spatial — tracked bounding boxes + occlusion levels (none / partial / full), object types (pedestrian, vehicle, traffic light, crosswalk, bus station), and 17-keypoint 2D pose.
2. Behavioral (frame-level, 6 attributes per pedestrian per frame) — crossing behavior, traffic interaction, pedestrian activity, attention indicators, social dynamics, stationary behavior.
3. Scene — intersection type, signalized type, road type, location type, motion direction, time of day.
4. Interaction — pedestrian–ego-vehicle interaction flag.
5. Location — spatial context (near divider, side of the road, near crosswalk, ...).

Additional pedestrian attributes: age, gender, carrying object, crossing motive, crosswalk usage, and whether the pedestrian crosses in front of the ego-vehicle.

See the GitHub README for the complete attribute taxonomy and per-attribute value definitions.

Usage

The dataset interface (iddped_interface_traj.py) in the GitHub repo consumes these annotations. Place this repo's contents under data/IDDPedestrian/ and follow the setup instructions in the repo to extract frames and run the baselines.

Citation

@inproceedings{idd2025ped,
  author    = {Ruthvik Bokkasam and Shankar Gangisetty and A. H. Abdul Hafez and C. V. Jawahar},
  title     = {Pedestrian Intention and Trajectory Prediction in Unstructured Traffic Using IDD-PeD},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  publisher = {IEEE},
  year      = {2025}
}