Pretraining & data augmentation boost drone detection accuracy to 150m
New DNN detects drones up to 150m away with 2x better out-of-domain performance
Deep Dive
Researchers developed a compact DNN for acoustic drone detection that uses pretraining on broad sound events (before fine-tuning on drone recordings) and on-the-fly augmentations (pitch shift, noise mixing, microphone simulation). The model achieves effective detection up to 150m on the IDMT Berne 2022 dataset and shows substantial true-positive rate improvements on out-of-domain data (AuDroK) compared to training from scratch, with low false positives on non-drone audio.
Key Points
- Pretraining on broad sound events (ESC-50) before drone fine-tuning is the dominant factor, doubling TPR on out-of-domain data.
- On-the-fly augmentations (pitch shift, noise mixing, mic simulation, spectrogram) further boost performance on the AuDroK benchmark.
- Effective detection at distances up to 150m on IDMT Berne 2022, with low false positives on traffic and environmental noise corpora.
Why It Matters
Enables low-cost, passive drone surveillance that works reliably across unseen environments—critical for airspace security and anti-drone systems.