ASPED: An Audio Dataset for Detecting Pedestrians

This repo contains the data for the ASPED dataset, presented at ICASSP 2024.

• Paper Link, Project Homepage

• Pavan Seshadri, Chaeyeon Han, Bon-Woo Koo, Noah Posner, Suhbrajit Guhathakurta, Alexander Lerch

Usage

This dataset contains audio and video recordings of pedestrian activity collected at various locations in and around Georgia Tech.

Labels of pedestrian counts per each second of audio/video are provided as well, calculated via a computer vision model (Mask2Former trained on msft-coco) using the video recordings.

Access

It is recommended to use the huggingface_hub library to download the dataset from this location. Info on downloading with huggingface_hub.

Downloading the entire dataset can be done with the following code:

from huggingface_hub import snapshot_download
snapshot_download(repo_id="pseshadri9/ASPED", repo_type="dataset")

Alternatively if you would like to download only the audio or video, pass the ignore_patterns flag to snapshot_download to avoid downloading the entire set.

Audio Only

from huggingface_hub import snapshot_download
snapshot_download(repo_id="pseshadri9/ASPED", repo_type="dataset", ignore_patterns="*.mp4")

Video Only

from huggingface_hub import snapshot_download
snapshot_download(repo_id="pseshadri9/ASPED", repo_type="dataset", ignore_patterns="*.flac")

Citation

@inproceedings{Seshadri24,
  title={ASPED: An Audio Dataset for Detecting Pedestrians},
  author={Seshadri, Pavan and Han, Chaeyeon and Koo, Bon-Woo and Posner, Noah and Guhathakurta, Suhbrajit and Lerch, Alexander},
  booktitle={Proc. of ICASSP 2024},
  pages={1--5},
  year={2024},
  organization={IEEE}
}