Neural Drone Localization Exploiting Signal Synthesis of Real-World Audio Data

As unmanned aerial vehicles (UAVs) become increasingly common, concerns about security, privacy, and noise pollution have intensified. As a result, the need for accurate and efficient UAV localization and tracking has become critical for security operations and timely intervention, yet algorithmic audio-based localization methods have limitations in complex outdoor environments. This study presents an approach to synthesizing drone acoustic signals and generating training datasets designed for deep neural network (DNN)-based localization. Using these simulated signals, two neural networks, SELDnet ACCDOA and Neural SRP, were trained and evaluated for accurate direction-of-arrival (DOA) estimation, addressing challenges specific to outdoor acoustic conditions. Their performance was benchmarked against the steered response power with phase transform (SRP-PHAT) methods. To further validate the models' effectiveness, real-world drone data were collected and used for testing. Experimental results indicate that neural networks trained on synthesized data achieve effectiveness comparable to SRP-PHAT, validating the reliability of the simulation approach, with Neural SRP even outperforming SRP-PHAT-based algorithms in DOA estimation accuracy.