Skip to content

Python tool for aircraft Doppler audio geolocation using historical ADS-B flight data (Source-available)

4.1 KB
Raw Blame

Checkpoint: 2026-11-02

The implementation campaign is active with five swarm workers and the primary integration agent. The task plan and GitHub issue #1 track current ownership. Publishing the checkpoint does not mark the automatic-upload objective complete. Shared contracts and whole-recording intake have landed; the remaining automatic stages are being implemented and integrated.

Acceptance target

Upload an MP3, discover its usable Doppler events automatically, match historical aircraft trajectories, and return a supported recording-location estimate. The user accepts approximately 100 m error. Missing evidence must produce an explicit inconclusive result rather than an unsupported precise coordinate.

Implemented

  • Audio/video decoding through ffmpeg; spectrogram contrast, bounded-slope Viterbi ridge extraction, straight-pass fitting, and noise/detection checks.
  • CSV, Parquet, KML and readsb trajectory import; type classification, geometric altitude handling with labelled fallbacks, and track-gap splitting.
  • WGS84 ECEF geometry, cubic Hermite trajectories, emission-time sound propagation, actual radial-velocity Doppler modelling, and no extrapolation.
  • Terrain/geoid raster sampling, cached Copernicus GLO-90 coverage, corridor grids, sparse CPA search, event-offset and silence filters, locked multi-event scoring, refinement, and inspectable JSON/CSV/NPZ/PNG outputs.
  • Download/ingestion scripts, an independent synthetic audio generator, and a 40-recording real feature benchmark.

Evidence

Artifact directory Finding
results/synthetic-two-events/ Correct pair, 21.8855 m position error, -0.3426 s clock error
results/synthetic-known-minute/ Correct flight, 40.8744 m position error
results/reference-local/ PC12 a1dfeb:1 + B350 ace391:3; 43.8585 m from mapped public-platform coordinate
results/reference-local-fast/ Same winning position and score as the preceding run; 46.26 s versus 103.09 s recorded runtime
results/reference-faint-alltypes-local/ CRJ9 a38237; 357.47 m error; fails the 100 m target
results/real-feature-benchmark/ 33/40 detections, zero processing errors; detected-case median absolute speed/slant errors 6.85%/15.37%

The real two-event winning point is 40.8769317589, -73.5249822950. The comparison point is the public nest-platform listing at 40.8766, -73.5247, whose precision and physical accuracy are limited. The search used a regional box informed by the source video. It is not a blind country-wide localization result.

The inferred recording start is 2026-09-08 20:17:29.662 UTC. The faint-event configuration rounds that result to 20:17:30 UTC; it is not an independently known recording clock. The video's stated 20:37 UTC and its various displayed timelines have not been fully reconciled.

The feature benchmark uses supplied harmonic-band annotations and contains substantial outliers. It validates acoustic measurements only. The synthetic renderer does not call the search propagation model, and its truth is read only for evaluation after search.

Fresh publication check: 11 tests passed, covering analytic retarded-time physics, full Unix timestamps, no extrapolation, repeated approaches, sparse/dense CPA agreement, filters, shared event timing, noise rejection, altitude decoding, and terrain/geoid sampling. These tests do not establish automatic discovery or national-scale localization.

Remaining work

  1. Automatic full-recording event discovery and harmonic grouping, replacing manual event windows and frequency bands.
  2. An MP3 upload interface and orchestration, including recording-date/time-window input when metadata do not establish it.
  3. Reliable candidate selection under curved flight paths, source-frequency drift, low SNR, altitude uncertainty and missing flight data.
  4. Scalable national search and independent real tests with withheld location and independently established recording times.
  5. Confidence/ambiguity reporting and negative controls.

See HANDOFF.md for suggested separable work and exact replay commands.