OpenFieldSensorDeployment/open-field-acoustic-readings

Open-Field Acoustic Readings Dataset

Distribution Statement A. Approved for public release; distribution is unlimited.

Dataset Summary

This repository contains synchronized outdoor acoustic recordings, vehicle GPS, sensor locations, processed 200 ms observations, and portable derived metadata for the single approved open-field vehicle-tracking sequence used by the associated experiment. It is not a release of other recordings collected at the deployment site. This repository contains dataset files and dataset documentation only. Research algorithms, experiment drivers, training code, plotting code, notebooks, and manuscript sources are not included.

The sequence uses nine active acoustic nodes: 1, 2, 3, 4, 5, 7, 8, 9, 10. Node 6 is retained in the source sensor-location table for provenance but was excluded before constructing the associated experimental instance.

Contents

Path	Description
data/audio/	Nine lossless multichannel FLAC recordings, one per active node.
data/gps/	The raw vehicle trajectory for sequence 20250812_1050.
sensor_locations/	Source and active-node sensor coordinate tables.
data/processed/	Aligned acoustic power/distance tables and valid row indices.
metadata/	Active configuration, portable learned embeddings, and replay/mixing metadata.
MANIFEST.csv	Per-file role, size, source basename, and SHA-256 digest.
CHECKSUMS_SHA256.txt	Checksums in standard verification format.

There are no examples_with_embed.pkl or bandit-instance pickle caches in this release. Those objects are implementation caches reconstructible from the released tables, arrays, metadata, and an experiment implementation.

Distribution and Usage Status

This dataset was approved for public release under:

Distribution Statement A. Approved for public release; distribution is unlimited.

Data Provider / Public-Release Approval

The released sequence was reviewed and approved by the relevant data provider for public release.

Sensitive Content and Privacy

The data come from an instrumented vehicle-tracking experiment. No human-subject data were intentionally collected. The release contains raw acoustic recordings and precise vehicle and sensor GPS coordinates. The released recordings were reviewed for incidental intelligible speech and personally identifiable or sensitive content, and none was identified.

No separate standard open-source or Creative Commons license is asserted by this dataset card. See LICENSE_OR_USAGE_NOTES.md for the approved usage notes.

Raw Formats

Audio

Each vehicle_audio_..._node_NN_respeaker.flac file is FLAC, 16 kHz, six-channel, signed 16-bit PCM. The files are lossless byte copies of the approved source recordings under neutral public filenames. Audio starts at the timestamp encoded in each filename. The associated preprocessing uses 10 ms mean-power chunks and resamples them into 200 ms observations.

Vehicle GPS

vehicle_gps_20250812_105042.csv is a headerless 11-field trajectory with 7,599 rows. Field 0 is local wall-clock time, fields 1 and 2 are latitude and longitude in decimal degrees (WGS 84 / EPSG:4326), and the remaining receiver fields are preserved verbatim. The associated preprocessing consumes the first three fields, removes duplicate timestamps, sorts chronologically, and aligns the trajectory with 200 ms audio observations. For exact compatibility, note that the original preprocessing called pandas.read_csv with its default header handling, so the first raw record supplied column labels and the resulting aligned table has 7,598 rows. Timestamps carry no encoded UTC offset; interpret them in the acquisition clock used by the synchronized sensor files.

Sensor Locations

The sensor CSVs contain Node #, Lat, Lon, and Acc. Latitude and longitude are WGS 84 decimal degrees. Acc is a source text field. During preprocessing, vehicle and sensor coordinates are projected to local UTM zone 18N (EPSG:32618), so distance_to_N values are in meters.

Processed Tables

gdf_cleaned.parquet has 7,598 timestamped rows and 18 columns: nine rpiN acoustic power columns in dB and nine distance_to_N columns in meters.

normalized_cleaned.parquet has the same 7,598-row datetime index and adds the aligned Latitude and Longitude columns, for 20 columns total. The name is retained for compatibility with the original processing output.

valid_indices.csv contains 7,473 zero-based row positions. A row is valid when it follows the preceding row by exactly 200 ms and projected vehicle movement is at least 0.0416667 m; the downstream context builder then uses five historical observations. The released cleaned tables contain no missing values. Use valid_indices.csv for the associated experimental sample set.

from pathlib import Path

import numpy as np
import pandas as pd

root = Path(".")
processed = pd.read_parquet(root / "data/processed/gdf_cleaned.parquet")
valid = pd.read_csv(root / "data/processed/valid_indices.csv")["valid_index"]
embeddings = np.load(
    root / "metadata/final_embedding_divisor_hidden_512/learned_embeddings.npy",
    allow_pickle=False,
)

Derived Metadata

learned_embeddings.npy is a portable float32 array with shape (964017, 24). Rows follow chronological valid-time order and, within each time, action order by subset size and canonical subset string. There are 129 actions per time. Keeping this array makes the learned representation stable across hardware; direct checkpoint inference was independently observed to agree within 7.2e-7 maximum absolute error.

The replay metadata records a circular modular-stride sequence with stride 25 (5 seconds at 200 ms sampling), a 7,208-step period, 20 clustered states, and the fitted mixing coefficient used by the associated experiment. See provenance_public.md for the complete dependency chain.

Filename Compatibility

Public filenames are intentionally neutral. MANIFEST.csv records each raw file's original source_basename. Legacy loaders can construct temporary aliases from that column without modifying this dataset. Such aliases are an implementation compatibility detail and are not included here.

Integrity And Limitations

Use CHECKSUMS_SHA256.txt to verify every released artifact. The release contains one vehicle pass at one outdoor site and should not be treated as a general benchmark for all vehicles, weather, terrain, microphones, or acoustic conditions. GPS accuracy, multipath, clock alignment, and node-specific audio conditions may affect downstream analyses. The nine recordings are multichannel; channel selection and aggregation are analysis choices.

Acknowledgement

This research reported in this paper was sponsored in part by: the IoBT Collaborative Research Alliance funded by the Army Research Laboratory (ARL) under Cooperative Agreement W911NF-17-2-0196. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the funding agencies.