13:30
Session 20: Acoustic Detection and Identification of Drones
Chair: Furkat Yunus
13:30
20 mins
|
Comparative Evaluation of Models and Features for Acoustic Drone Detection in Real-World Conditions
Martin Blass, Hannes Bradl, Franz Graf
Abstract: Acoustic sensors offer a promising approach for detecting unmanned aerial vehicles (UAVs), complementing visual or radio-frequency-based methods and remaining effective in scenarios where other sensing technologies are limited. However, reliable acoustic drone detection in real-world environments remains challenging due to varying background noise, recording conditions, target distances and UAV operating states. This contribution investigates single-channel acoustic drone detection formulated as a binary classification task, using a large set of manually annotated field recordings collected under realistic outdoor conditions. The study compares several detection approaches, ranging from classical machine learning classifiers using manually engineered acoustic features to deep learning architectures operating on log-Mel spectrogram inputs. For the deep learning models, different training strategies are evaluated, including training from scratch, pre-trained frozen feature extractors and fine-tuning networks. In addition to frame-based classification performance, the models are assessed with respect to computational complexity and CPU-based inference runtime. The results show that machine learning models with engineered features remain highly competitive, achieving strong detection performance at lower computational cost. Fine-tuning improves the performance of pre-trained deep models compared to frozen configurations and training from scratch, indicating that adaptation to the target domain is important for real-world UAV detection. Overall, the findings highlight a trade-off between detection performance and computational efficiency, with feature-based models providing attractive solutions for resource-constrained deployments and fine-tuned deep models offering the highest modeling capacity where sufficient computational resources are available.
|
13:50
20 mins
|
Performance comparison of MEMS and analogue microphones for acoustic drone localization
Nathan Itare, Jean-Hugh Thomas, Stephane Letourneur, Kosai Raoof
Abstract: The use of microphone arrays for source localization is a valuable tool to deal with threats caused by malicious or invasive drones. The localization performance depends on the signal processing techniques, but also on the design of the microphone arrays. Several factors are influential in the design: the number of microphones, the microphones’ disposition and their type. This study focuses on the influence of the type of microphone used in an array. Two types of microphones are investigated, yielding two different arrays. The first array is composed of MEMS microphones, which are compact and low-cost but have limitations in terms of sensitivity, and the second array uses analogue microphones, which are more precise, but more expensive. One constraint of the study is the number of microphones: only 10 microphones are employed in each array to perform the localization. Two elements are used to compare the arrays: the spectral content of the measured signals, and the accuracy of the direction of arrival. In the article, the signal processing needed to enhance the performance of the localization with MEMS is described, as well as the localization method used with both types of arrays. The localization method is based on delay and sum beamforming with a consideration of the drone acoustic signature. The comparison is performed with outdoor experimental measurements using a DJI Matrice 4T.
|
14:10
20 mins
|
Improving acoustic drone detection generalization through pretraining and data augmentation
Paul Reuter, Mattes Ohlenbusch, Christian Rollwage
Abstract: Detecting unauthorized UAV flights is critical for surveillance, security, and airspace management. Acoustic drone detection, which relies on the distinctive propeller and motor sounds of UAVs, provides a low-cost, passive solution that requires no line of sight. A central challenge is generalization: reliably distinguishing drone signatures from ambient noise across unseen recording setups, environments, and UAV types (out-of-domain). Inspired by advances in large-scale audio pretraining, we develop a compact DNN-based detector and improve its generalization by (1) pretraining the model for broad sound-event classification before fine-tuning on diverse in-house and public drone recordings, and (2) applying on-the-fly augmentations (pitch shifting, noise mixing, microphone transfer function simulation, spectrogram augmentation) to expose the model to varied acoustic conditions. An ablation study quantifies the impact of each augmentation. For evaluation, we set target false-positive rates (FPR) aligned with real-world surveillance needs and report true-positive rates (TPR) on both in-domain data (public IDMT Berne 2022) and out-of-domain data (public AuDroK). Our results show that pretraining is the dominant factor for robust detection, yielding substantial TPR improvements over training from scratch on all benchmarks. The full augmentation chain provides additional gains on acoustically mismatched out-of-domain data, achieving the best mean TPR on the AuDroK subsets and the largest improvements on the most challenging scenarios. We further validate real-world applicability by measuring false positives on public non-drone corpora (IDMT-TRAFFIC and ESC-50), demonstrating equally low FPR on unfamiliar backgrounds. A distance-dependent analysis on IDMT Berne 2022 shows effective detection at distances up to 150 m.
|
14:30
20 mins
|
Implementation and Optimization of a Compact Deep Learning Architecture for Portable Acoustic Drone Detection
Hadrien Pujol, Julien Preuilh, Magali Arnaud, Thierry Mazoyer
Abstract: Counter-drone defense is becoming an increasingly critical topic within the defense industry. Acoem, through its subsidiary Metravib Defence—a long-standing industrial leader in acoustic threat detection and localization—has launched an intensive R&D program over the past few years. This program focuses on deep learning-empowered algorithms dedicated to drone detection, classification, and localization. The objective is to embed real-time processing onto miniaturized, low-power, fully passive acoustic sensors for operational use.
The work presented in this paper describes the optimization of a CNN, derived from previous BeamLearning-ID research, to fit within a specialized Neural Processing Unit (NPU). The goal is to detect drones with high false-alarm rejection and rapid reaction times. Beyond the CNN detector, the overall processing chain includes a time stabilizer and a high-resolution localization step, leveraging the system prototype's multichannel miniature acoustic array.
After a brief system overview, this paper focuses on CNN design and optimization with an emphasis on model transparency. We aim to draw parallels between the filter coefficients of trained convolutional layers and the underlying physics of signal processing. This approach helps mitigate the "black-box" effect often criticized in deep learning. Drawing on field tests involving over 20 hours of acoustic flight signals from various drone models and attack scenarios, we present a physically-guided optimization of the CNN. This ensures maximum performance while maintaining the compact network size required for real-time inference on embedded systems.
Finally, the CNN's ROC curves are analyzed to evaluate detection rates and false alarms, with comparisons to existing literature. We also demonstrate how time-stabilizing functions enhance alert robustness. The paper concludes with technical and operational perspectives on the future of AI-empowered acoustic detectors.
This work was performed using HPC resources from GENCI-IDRIS (Grant 2025-AD011013877R2])
|
14:50
20 mins
|
AeroFeathers: Biomimetic 3D Printed Fiber Surface Treatments for Quiet Drone Propellers
William Johnston, Bhisham Sharma, Janith Godakawela, Kara Hardy
Abstract: The acoustic signature of small unmanned aerial vehicle (UAV) propellers remains a persistent challenge for operation in noise-sensitive environments and contributes to drone detectability through acoustic sensing. Previous bio-inspired approaches have explored rigid serrations and geometric blade modifications, but few replicate the compliant fibrous structures associated with the silent flight of owls. We investigated flexible fiber-based surface treatments fabricated directly onto drone propeller blades using fused deposition modeling (FDM). A proprietary toolpath modification algorithm, operating directly on sliced propeller geometries, generates three controllable fibrous structures—leading-edge serrations, a downy coat, and trailing-edge fringes—with prescribed fiber separation, thickness, and length. Fourteen bio-inspired propeller configurations incorporating isolated and combined features were fabricated and tested with simultaneous acoustic and aerodynamic measurements. One combined-feature configuration (downy coat with trailing-edge fringes) achieved a reduction of approximately 2 dB in near-field overall A-weighted sound pressure level (OASPL) at 50 gf thrust with nearly the same reported mean mechanical power as the printed clean baseline. Several other configurations achieved acoustic reductions but at the cost of substantially higher power consumption. Although preliminary, these findings suggest that additively manufactured fibrous surface treatments may provide a manufacturable pathway toward quieter UAV propellers.
|
|