Datasets:
Dataset Card for EWFootstep 1.0
Dataset Summary
EWFootstep 1.0 is a real-world footstep audio dataset collected from 176 subjects across three geographic regions of India (Central Region — plateau terrain; Eastern Region — dense jungle; Northern Region — rocky hills and jungle). It was built to support early-warning (EW) and surveillance systems, audio forensics, and domain adaptation research in acoustics, by capturing footstep acoustic signatures that distinguish:
- Environment:
forest,road,indoor - Occupancy:
single person,multiple persons
Unlike prior footstep datasets, which are typically collected in controlled lab settings for person identification or biometrics, EWFootstep 1.0 is recorded entirely in the field, under authentic, uncontrolled acoustic conditions, making it directly usable for training and evaluating models intended for real-world deployment (e.g., counter-insurgency/counter-terrorism early-warning systems, remote border/checkpoint monitoring, crime-scene audio forensics).
Note on this release: base field recordings were augmented (codec, reverb, low-pass filter, pitch shift, time stretch, volume variation) to increase training data volume. All 6
environment × num_personsclasses are populated in this upload (799 clips total across 139 base recordings). Splits and CV folds are assigned at the base-recording level, so every augmented variant of a given clip stays together in the same split/fold — see Data Splits.
The dataset has been used as the benchmark in two peer-reviewed papers (see Citation):
- A dataset paper introducing EWFootstep 1.0 and validating it with time/frequency-domain analysis, Fréchet Audio Distance (FAD), t-SNE, Davies–Bouldin Index (DBI), and ML/CNN baselines.
- A follow-up paper proposing a Hierarchical Multi-Task Learning (H-MTL) model (CNN + Transformer) that achieves 96.20% accuracy (environment) and 96.60% accuracy (number of persons) on this dataset.
Supported Tasks and Leaderboards
- Environmental Sound / Acoustic Scene Classification: classify recordings into
forest/road/indoor. - Occupancy / Person-count classification: classify recordings into
single person/multiple persons. - Hierarchical / Multi-task classification: joint prediction of environment (main task) and number of persons (auxiliary task), as in the H-MTL model from the Scientific Reports paper.
- Audio forensics research: environment and crowd-size inference from ambient/footstep audio for crime-scene reconstruction.
- Domain adaptation in acoustics: the dataset spans multiple regions/terrains and recording devices, enabling cross-domain generalization studies.
No official leaderboard currently exists; reported baselines are in Additional Information.
Languages
Not applicable (non-speech acoustic/environmental audio dataset).
Dataset Structure
Data Instances
Each instance is a .wav audio clip of footstep sounds, resampled to 44.1 kHz, 16-bit PCM, renamed to <environment>-<num_persons>-<n>.wav (e.g. indoor-single-1.wav) for consistency; the original raw filename is preserved in the source_file field. A conceptual instance:
{
"audio": "forest/multiple/forest-multiple-133.wav",
"environment": "forest",
"num_persons": "multiple",
"kfold": 1, // -1 for test rows
"source_file": "double 8.wav",
"sample_rate": 44100,
"bit_depth": 16
}
Data Fields
| Field | Type | Description |
|---|---|---|
audio |
Audio (wav) |
Footstep recording, 44.1 kHz / 16-bit PCM |
environment |
ClassLabel |
One of forest, road, indoor |
num_persons |
ClassLabel |
One of single, multiple |
kfold |
int |
Fold id 0-4 for 5-fold cross-validation, assigned over the combined train+validation pool only; -1 for test rows (always held out, not part of any fold) |
source_file |
string |
Original raw filename before renaming (kept for traceability back to the source export) |
The composite classes formed by environment × num_persons mirror the hierarchical representation used in both source papers, and all six are populated in this release: forest-single, forest-multiple, road-single, road-multiple, indoor-single, indoor-multiple.
Data Splits
Augmentation-aware, leakage-free splitting: the source audio for each base recording includes up to 6 augmented variants (audio_codecs, audio_reverb, low_pass_filter, pitch_shifted, time_streched, volume_variation) plus the original clip. Splitting at the individual-file level would let augmented copies of the same underlying recording leak across train/validation/test (or across CV folds), inflating evaluation metrics. Instead, files are first grouped by base recording, and the 70/15/15 split (stratified per environment × num_persons class) is assigned at that group level — every variant of a given base recording always lands in the same split and the same CV fold. scripts/build_dataset.py implements this and is rerun whenever more audio is added.
| Split | Files | Base recordings |
|---|---|---|
| Train | 561 | 98 |
| Validation | 126 | 21 |
| Test | 112 | 20 |
| Total | 799 | 139 |
Per-class breakdown (files / base recordings):
| Class | Train | Validation | Test |
|---|---|---|---|
| forest-multiple | 147 / 21 | 28 / 4 | 34 / 5 |
| forest-single | 46 / 13 | 15 / 3 | 14 / 3 |
| indoor-multiple | 56 / 14 | 13 / 3 | 6 / 3 |
| indoor-single | 88 / 18 | 21 / 4 | 16 / 3 |
| road-multiple | 140 / 20 | 28 / 4 | 28 / 4 |
| road-single | 84 / 12 | 21 / 3 | 14 / 2 |
Because splitting happens at the base-recording level (to prevent leakage) rather than the file level, per-class file counts approximate but don't exactly hit 70/15/15 — some classes have as few as 12-21 base recordings, so validation/test slices for those classes are small; treat per-class test metrics on road-single and indoor-multiple with appropriate caution.
Cross-validation: a kfold column (0-4) is assigned over the combined train+validation pool, at the same base-recording level and stratified per class, so users can run 5-fold CV for model selection/tuning without touching the held-out test split. To reconstruct the CV pool:
from datasets import load_dataset, concatenate_datasets
ds = load_dataset("happyman11/EWFootstep1.0")
cv_pool = concatenate_datasets([ds["train"], ds["validation"]])
fold_0_val = cv_pool.filter(lambda x: x["kfold"] == 0)
fold_0_train = cv_pool.filter(lambda x: x["kfold"] != 0)
(Working from a local clone of this repo instead of the Hub? Use load_dataset("audiofolder", data_dir="data") from the repo root.)
Dataset Creation
Curation Rationale
Existing footstep datasets (e.g., Algermissen et al.'s 1,125-sample/5-subject set, PURE, AFPILD, UPC-TALP, ESC-50, FSD50K) are collected in controlled lab environments, are small, and/or focus on person identification rather than environment or occupancy classification. EWFootstep 1.0 was created to fill this gap by providing a large, field-collected, hierarchically labeled dataset suitable for early-warning and forensic applications where visual surveillance is impractical (poor lighting, occlusion, remote/administratively-restricted areas).
Source Data
Data Collection Process
- Recorded across three regions of India: Central (plateau), Eastern (dense jungle), Northern (rocky hills/jungle).
- Microphones: AKG C214 omnidirectional (150 dB sensitivity, SNR > 70 dB), placed on the ground 2–4 ft laterally from the walking track; iPhone 12 MEMS mic used where a standard mic could not administratively be set up (hilly terrain).
- Walking track: 30–40 ft between start and finish points; subjects walked naturally to preserve natural cadence variation.
- Recorded continuously in batches of 10–15 subjects (mic could not practically be started/stopped between individuals in field terrain), then segmented into individual samples.
- Footwear varied deliberately across soft-soled (sports shoes) and hard-soled (leather boots) to capture acoustic variation from rubber/leather/plastic soles.
- Original recordings captured as
.mp4and converted to.wavvia Python'spydub. - Resampled to 44.1 kHz / 16-bit PCM (Nyquist 22.05 kHz, well above the <2.5 kHz dominant footstep energy band).
Who are the source subjects? 176 individual subjects, aged 18–45 years, heights 163–175 cm. Informed consent was obtained from all subjects for dataset publication. The study was approved by the Institutional Ethics Committee of the Indian Institute of Technology Indore.
Annotations
Labels (environment and occupancy) were assigned at collection time based on the controlled recording protocol (known environment, known number/identity of walkers), rather than via post-hoc crowd annotation.
Personal and Sensitive Information
Recordings capture footstep/gait acoustic signatures, which some literature treats as a soft-biometric identifier (see Algermissen et al., Altaf et al. cited in the source papers). No speech, names, or other directly identifying information is intended to be included. Consent for publication was obtained from all subjects. Downstream users should not attempt to re-identify individuals from gait acoustics.
Considerations for Using the Data
Social Impact and Biases
This dataset was designed explicitly for security, surveillance, and forensic use cases, including counter-insurgency/counter-terrorism (CI/CT) early-warning systems, as described in the companion Scientific Reports paper. Users and downstream deployers should consider:
- Dual-use risk: acoustic surveillance capability could be repurposed for unauthorized monitoring or tracking of individuals; deployments should comply with local law and privacy norms.
- Geographic/demographic scope: subjects are limited to specific regions of India, age range 18–45, and a specific height range — models trained solely on this data may not generalize to other populations, footwear cultures, or terrains (e.g., snow, urban pavement outside India, desert).
- Class imbalance: file counts per class range from 75 (
forest-single,indoor-multiple) to 209 (forest-multiple) — see the per-class table in Data Splits and account for this during training (e.g. class weighting or balanced sampling).
Other Known Limitations
- Environment labels reflect the recording site, not a general-purpose acoustic scene taxonomy — models may not transfer directly to unrelated ASC benchmarks (ESC-50, FSD50K, etc.).
- Multi-person recordings do not disambiguate exact group size (2 vs. 3+) — only a binary single/multiple label is provided in v1.0.
- Field recording conditions mean microphone-to-subject alignment isn't perfectly centered in every clip (noted explicitly by the original authors).
Additional Information
Reported Baselines (from the source papers)
| Model | Task | Test Accuracy | Test F1 |
|---|---|---|---|
| SVM | Environment | ~65–71% | comparable |
| KNN | Environment | ~65–71% | comparable |
| CNN | Multi-class (6-way) | 88% | 87% |
| CNN + Attention | Multi-class (6-way) | 92% | 92% |
| MTL | Environment (T1) | 87.4% | — |
| MTL | Num. persons (T2) | 75.6% | — |
| H-MTL (proposed, w/ PE) | Environment (main task) | 96.2% | 96.4% |
| H-MTL (proposed, w/ PE) | Num. persons (auxiliary task) | 96.6% | 96.7% |
| Human baseline | Both tasks | ~66–71% (25–30 pts below H-MTL) | — |
Clustering quality (Davies–Bouldin Index, lower is better): raw features 9.5–11.8 → CNN features 0.62 → CNN+Attention features 0.37.
Baselines above are from the full multi-region dataset described in the source papers, not solely the indoor-single subset in this release.
Dataset Curators
Anshuman Agrahri, Chandresh Kumar Maurya (IIT Indore); Ravi Shekhar Tiwari (IIIT Delhi); Shashwat Tripathi (University of Glasgow).
Licensing Information
The two associated papers are published open access under CC BY-NC-ND 4.0 (non-commercial use, sharing, and reproduction permitted with attribution; no derivatives; no permission to share adapted material). This card defaults to cc-by-nc-nd-4.0 to mirror the papers — update it if the data-sharing agreement for this specific upload specifies otherwise.
Citation Information
If you use this dataset, please cite both papers:
@article{agrahri2026ewfootstep,
title = {EWFootstep 1.0: footstep data for early warning and surveillance systems},
author = {Agrahri, Anshuman and Maurya, Chandresh Kumar and Tiwari, Ravi Shekhar and Tripathi, Shashwat},
journal = {EURASIP Journal on Audio, Speech, and Music Processing},
volume = {2026},
number = {1},
pages = {32},
year = {2026},
publisher = {Springer},
doi = {10.1186/s13636-026-00465-2},
url = {https://doi.org/10.1186/s13636-026-00465-2}
}
@article{agrahri2026acoustic,
title = {Acoustic sentinel: hierarchical classification of footstep sound using fine and coarse-grain acoustic feature representations for tactical surveillance},
author = {Agrahri, Anshuman and Maurya, Chandresh Kumar and Tiwari, Ravi Shekhar and Tripathi, Shashwat},
journal = {Scientific Reports},
volume = {16},
pages = {5635},
year = {2026},
publisher = {Nature Publishing Group},
doi = {10.1038/s41598-026-35756-3},
url = {https://doi.org/10.1038/s41598-026-35756-3}
}
Contributions
Dataset card compiled from the two publications above for Hugging Face Hub release.
- Downloads last month
- 79