Acoustic Imaging for UAV Detection: Dense Beamformed Energy Maps and U-Net SELD
Summary: arXiv:2508.00307v3 Announce Type: replace-cross
The detection of Unmanned Aerial Vehicles (UAVs) has become a pressing concern in various fields, including security, environmental monitoring, and urban planning. Recent advancements in acoustic imaging techniques present a novel approach to UAV detection, leveraging dense beamformed energy maps and U-Net models for effective sound source localization.
Introduction
In the study presented in the paper, researchers introduce an innovative U-Net model specifically designed for 360-degree acoustic source localization. This approach is formulated as a spherical semantic segmentation task, which significantly enhances the accuracy and reliability of UAV detection.
Methodology
Unlike traditional methods that focus on regressing discrete direction-of-arrival (DoA) angles, this model segments beamformed audio maps into distinct regions indicative of active sound presence. The methodology involves several key steps:
- Data Collection: A custom 24-microphone array is employed to capture audio signals, which are synchronized with drone GPS telemetry. This process allows for the generation of binary supervision masks that are essential for training the model.
- Beamforming Technique: The researchers utilize delay-and-sum (DAS) beamforming to enhance the quality of the audio maps, thereby improving the accuracy of localization.
- U-Net Architecture: A modified U-Net architecture is trained on frequency-domain representations of the beamformed energy maps. The model learns to identify spatially distributed source regions while addressing class imbalance through the implementation of the Tversky loss function.
Results and Applications
The experimental results demonstrate that the U-Net model generalizes effectively across various environments. The approach not only provides improved angular precision but also establishes a new paradigm for dense spatial audio understanding that extends beyond traditional Sound Source Localization (SSL) techniques.
Moreover, the dataset utilized in the study comprises real-world open-field recordings of a DJI Air 3 drone, accompanied by synchronized 360-degree video and flight logs from multiple dates and locations. This comprehensive dataset plays a crucial role in training and validating the model’s performance.
Generalization and Future Work
In addition to UAV detection, the researchers validated the beamforming-plus-segmentation formulation on the DCASE 2019 TAU Spatial Sound Events benchmark. This validation indicates that the proposed method can effectively generalize to multiclass Sound Event Localization and Detection (SELD) scenarios, showcasing its versatility and potential for broader applications.
Conclusion
The introduction of a U-Net model for acoustic imaging in UAV detection represents a significant advancement in the field of sound source localization. By operating on beamformed energy maps and leveraging a robust training methodology, this approach paves the way for future research and applications in various domains, including environmental monitoring, security, and transportation.