Skip to content

v0.1.0 - The Acoustic Drone Detection Suite comes to life!

Latest

Choose a tag to compare

@JoulesSpace JoulesSpace released this 02 Jun 00:52
· 3 commits to main since this release

A Rust, edge-to-server, multi-task acoustic drone-detection toolbox with automated, honesty-first benchmarking. Docker-first, AGPL.

✨ AI summary below

Architecture

  • 15 crates over a shared no_std DSP backbone (drone-dsp: windowing, real FFT, spectral features) + a pluggable eval
    harness (drone-bench).
  • Insightface-style: one backbone, many task heads, common JSON-emitting benchmark.

Detection (drone-bench)

  • 14 pluggable approaches: band-ratio, HPS, spectral-gate, template, spectrogram-template, cepstrum, envelope, MFCC-LR,
    GTCC-LR, MFCC-MLP, feature-fusion, fusion-ensemble, physics-fused, sentry.
  • Metrics: F1, calibrated-F1, ROC/PR-AUC, Brier, real-time factor; ROC/PR/cost-quality plots.

Other task heads

  • drone-doa - GCC-PHAT direction-of-arrival (ULA, no_std core, simulated).
  • drone-id / drone-vendor - drone type (Al-Emadi) and 32-brand recognition.
  • drone-freq - blade-pass frequency / RPM estimation.
  • drone-range - distance estimation (air-absorption-tilt cue).
  • drone-bss - FastICA blind source separation for multi-UAV / ego-noise.

Honest evaluation (the differentiator)

  • Cross-dataset + hard-negative eval (xeval), truly-held-out unseen-drone test (heldout32), sample-rate/bit-depth and
    SNR-robustness sweeps.
  • Head-to-head vs a faithful upstream mel-CNN (drone-cnn): perfect in-distribution, collapses on unseen drones - ours
    generalize materially better on the only trustworthy test. Leakage is measured, not hidden.

Deployment (MCU → phone → desktop → server)

  • drone-edge no_std detector cross-builds to riscv32imc; drone-firmware real esp32-C6 firmware (~17-27 KB flash).
  • drone-mobile C-ABI/JNI for Android/iOS; drone-live real-time mic capture + hardware probe + alert + labelled
    field-recording.
  • Hardware tiers + model cards.

Infra & process

  • Docker Compose services (detector/dev/bench/plot/data), one-command check suite (fmt, clippy -D warnings, tests, no_std
  • riscv builds, folder lint).
  • Prior-art/SOTA research, design decisions, and a data-dense signal-chain infographic; honest limitations documented
    (field validation is the next frontier).