A YouTube creator says he can determine where a video was recorded to within about 100 metres using nothing but the aircraft noise in its background — noise the person holding the camera cannot hear.
The channel, Colsto, has 105,000 subscribers. In a video published this week he took a surveillance camera clip and worked from almost nothing: the date, 8 September, and the knowledge that it was shot somewhere in the continental United States. He located the camera, he says, to within a few hundred feet, and named the place as Oyster Bay, New York.
The signal is the Doppler curve of a propeller aircraft passing overhead. As a plane approaches and recedes, the pitch of its drone bends; the bend is inaudible to a listener but shows up in a spectrogram once you zoom in and raise the resolution. Low frequencies pass through walls, so the trace appears in almost any recording with ambient sound — including, he demonstrated, audio that has been through telephone compression.
Everything after that is scale. He downloaded a full day of ADS-B flight data for the continental United States with a Python script, a job that took hours. ADS-B carries position, speed, altitude and aircraft type code; the type code let him separate jets from propeller planes, and aircraft he could not classify he handled by elimination, filtering out known jets. He also needed a national elevation model, because ADS-B reports altitude above sea level while sound propagation depends on slant distance to the microphone, and a 3D physical model of the propagation itself. He says the project ran to more than 5,000 lines of code over several months.
The clip contained two overflights 920 seconds apart. Step one: find every place where a listener would hear one aircraft and then, 920 seconds later, another — accounting for the time sound takes to arrive. Inside a test area covering less than 0.05% of the continental US, that filter still left about 35,000 candidate points.
Step two: derive each aircraft's speed and range from the frequency decay of its Doppler curve. Across the whole country that produced roughly 73,000 candidates, many near airports.
Step three used an absence. Between the two overflights there were about fifteen minutes with no aircraft audible at all. Any location where a plane would have been heard in that window — near an airport, say — was eliminated. About 15,000 points remained.
Step four: for each survivor, model the Doppler curve the actual flight data would have produced and score it against the spectrogram. The highest score by a wide margin was the real camera location, which he then checked against Street View — the tree line matched, and the camera was visible in the images. In a second test he recovered the same location from one minute of audio with a known recording time, even though the Doppler curve in that clip was barely legible.
The claims come with real caveats. He has published neither his code nor a tool, so others cannot check his work; several people have rebuilt the method on GitHub, but without the demo material the accuracy figure cannot be verified independently. Generalisation is also open — an earlier video of his located places from thunder at about 5 km accuracy, three orders of magnitude coarser.
He is also pessimistic about defences. High-pass filtering does not remove it; neither does drowning it out. What works, he says, is cutting the video in a way that destroys the interval between two aircraft, or adding noise suppression — both of which alter the recording rather than the room. He reckons the practical risk is low, because there are probably a hundred easier ways to find someone.
The uncomfortable version of that argument is that the hundred easier ways are what made this one cheap to ignore.
