Audio & Speech

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.