Acoustic Detection of Unmanned Aerial Vehicles Using Deep Learning and Smartphone Based Sensing
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Dataset: Smartphone Acoustic UAV Detection — Field Recordings, Annotations, Models, and Code Overview This dataset accompanies the paper "Acoustic Detection of Unmanned Aerial Vehicles Using Deep Learning and Smartphone-Based Sensing: A Field-Oriented Study." It provides annotations, derived features, trained model weights, and evaluation code for binary acoustic drone detection using commodity Android smartphones. Raw audio is not distributed to protect incidental privacy; instead, we provide per-second ground-truth labels, per-window labels (3-second windows, 1-second hop), log-Mel filterbank features (128 bands), and raw model prediction probabilities sufficient to independently reproduce all reported results. Dataset Summary 47 field recording sessions totaling 6.2 hours of monophonic audio at 16 kHz 4 drone platforms: DJI Matrice 200, DJI Matrice 300, DJI Mavic 3, DJI Phantom 4 4 Android smartphone models: Samsung S22+, Samsung A52M, OnePlus Nord, Xiaomi Redmi 12 Environments: open field and lightly urban outdoor scenes, distances 10–120 m 126 minutes drone-present, 157 minutes drone-absent (flight-derived), 90 minutes ambient hard negatives Session-level train/validation/test splits (36/4/7) to prevent data leakage Contents metadata/ — Master session manifest (sessions.csv), train/val/test split assignments (splits.json), human listener baseline latencies, log-Mel normalization statistics annotations/ — Per-session segment-level and window-level ground-truth labels; pilot annotation campaign data with inter-annotator agreement features/ — Log-Mel filterbank features (128 bands, 25 ms frame, 10 ms hop) for reproducibility without raw audio models/ — Trained BEATs fine-tuned detector (full precision and INT8-quantized variants) with model card documenting architecture, training configuration, and ethical considerations results/ — Per-window prediction probabilities for all three models (BEATs, compact CNN, MFCC+SVM), ROC/PR curve data, per-session confusion matrices code/ — Complete training, evaluation, and preprocessing pipeline: training script with config matching the paper, evaluation script reproducing all reported tables with bootstrap confidence intervals, temporal smoothing (Algorithm 1), leave-one-device-out evaluation, publication-quality figure generation, windowing, log-Mel extraction, and data augmentation Reproducing Paper Results pip install -r code/training/requirements.txt python code/evaluation/evaluate.py --predictions results/test_predictions.csv --threshold 0.6 --K 5 --bootstrap This reproduces Table 5 (overall performance), generates ROC/PR curves, and computes 95% bootstrap confidence intervals. Key Results (Table 5) Model Precision Recall F1 FPR ROC AUC MFCC+SVM 0.78 0.75 0.76 0.10 — Compact CNN 0.90 0.89 0.89 0.07 — BEATs (proposed) 0.93 0.90 0.92 0.04 0.983
Fields of Study & Subjects
Music and Audio Processing
Speech Recognition and Synthesis
Noise Effects and Management
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 28 March 2026
- Language
- EN
- Document type
- Dataset