- Python 100%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| code | ||
| dataset | ||
| CITATION.cff | ||
| Code Description | ||
| Data and Code Availability | ||
| dataset_description | ||
| How to Reproduce the Analysis | ||
| LICENSE | ||
| README.md | ||
| Reproducibility Notes | ||
| requirements.txt | ||
| Software Requirements | ||
NLOS_acoustic_identity_recognition
Zenodo DOI: https://doi.org/10.5281/zenodo.21802627
Topological–Spectral Fusion of Secondary Acoustic Echoes for Identity Recognition in Non-Line-of-Sight Environments
This repository contains the Python source code, detailed dataset metadata, representative anonymized multichannel acoustic recordings, and reproducibility documentation associated with the manuscript:
Topological–Spectral Fusion of Secondary Acoustic Echoes for Identity Recognition in Non-Line-of-Sight Environments
Study Overview
The study investigates person identity recognition in a controlled airborne non-line-of-sight (NLOS) environment using delayed multichannel acoustic responses. Measurements were acquired from 10 anonymous subjects with an active array consisting of eight speakers and eight synchronized microphones. The experimental design includes 10 object configurations, three postures, four orientations, eight sequential transmitter conditions, and one simultaneous multi-transmitter condition.
Raw recordings are divided into non-overlapping 1,024-sample windows. The final window exceeding an adaptive recording-specific energy threshold is selected as the secondary-echo segment. Time-averaged STFT features and LPVG degree-sequence features are extracted and combined through feature-level concatenation.
The repository includes the final implementations for:
- Random Forest, Extra Trees, and SGD-based linear SVM experiments;
- object-configuration-based leave-one-scene-out (LOSO) validation;
- independent and synchronized secondary-window selection ablation;
- deep-learning baselines using raw and STFT-based inputs;
- single-source cross-material transfer and leave-one-material-out analyses;
- matched human-present versus human-empty validation.
Software Requirements
Install the core dependencies with:
pip install -r requirements.txt
The workflow was developed in Python and uses packages including NumPy, pandas, SciPy, librosa, scikit-learn, matplotlib, NetworkX, joblib, and PyTorch for the deep-learning baselines. A suitable PyTorch build should be installed for the target CPU or GPU environment.
How to Reproduce the Analysis
The scripts are designed to be opened in Spyder and executed with F5. They may also be run from a terminal.
Set the full-data location before running an analysis:
NLOS_AUDIO_ROOT = path to the wall-reflection WAV dataset
NLOS_OUTPUT_ROOT = optional output directory
Main entry points:
code/run_classical_loso.py
code/run_synchronized_window_ablation.py
code/run_deep_learning_baselines.py
code/run_cross_material_analysis.py
code/run_human_empty_validation.py
Command-line example:
cd code
python run_classical_loso.py
Citation
Please cite the associated article and Zenodo record when using this repository. The citation metadata are also provided in CITATION.cff.
Olgun, N. (2026). NLOS Acoustic Identity Recognition: Source Code and
Reproducibility Materials, version 1.0.0. Zenodo.
https://doi.org/10.5281/zenodo.21802627