Two webcams located from aircraft sounds

Blind acoustic Doppler geolocation of two public live cameras, checked against their published locations · recordings 2026-11-02

Passing aircraft leave faint falling tones in outdoor audio. The shape of each tone depends on where the listener was. We recorded three hours from each of two public YouTube live cameras, found the flyovers in the audio automatically and matched them against ADS-B flight tracks. For each camera we searched everywhere within 40 km of a nearby city, without using the stream titles, video or any published location. Only afterwards did we compare the answers with where the cameras are known to be.

How it works

A propeller's tone drops in pitch as the aircraft passes, because of the Doppler effect. How fast and how far it drops depends on the aircraft's speed, height and distance from the listener. For every 50 m cell we took the real ADS-B track of each aircraft heard, computed the exact tone a listener there would have heard, allowing for the time the sound takes to arrive, and scored the match against the recording. Adding the scores of several flyovers leaves one place that fits them all. We also check silence: a place near which loud aircraft passed unheard is ruled out.

Ashland, Virginia: Virtual Railfan railway camera (PTZ)

The audio puts the microphone 43 m from the published camera location. We estimated 37.75930, -77.48153 with an uncertainty of about 500 m, from 9 flyovers. The camera is mounted on the station building, which is also the town's visitor centre.

Each coloured band shows where you could stand and hear one flyover's curve from the aircraft we matched it to; the colours match the table below. One flyover allows a whole band of places. The camera is where the bands cross, inside the white contours of the combined fit.

Dashed lines are the Doppler curves predicted from each aircraft's ADS-B track at the audio estimate, with only the source frequency fitted. f0 is that fitted source frequency: the tone the aircraft emits, before the Doppler shift. Fit is the match against the traced ridge, where 1 is perfect.

Silence helps too. At the audio estimate, 7 of the 9 propeller aircraft that passed within 3 km are heard. In central Richmond, for comparison, five passed within 3 km, including two helicopters within 1 km, and none of them is in the audio.

The whole search area

We scored every 200 m square inside the dashed circle without knowing where the camera was. Brighter means a better fit to all the flyovers together. Only one place fits them all. The best places anywhere else fit about a third as well or less, too faintly to show at this scale.

• the curves of five single flyovers peak 0.5 to 2 km north-east of the joint estimate, a modelling residual we could not remove
• three clear flyovers match no ADS-B track at any location

Audio: the live stream on YouTube, recorded with its HLS clock. Published location: Waymarking: livestream web camera at the train depot, Ashland · Town of Ashland: Virtual Railfan brings Ashland Train Day to the world; coordinates from OpenStreetMap.

Ithaca, New York: Cornell Lab FeederWatch bird-feeder cam

The audio puts the microphone 215 m from the published camera location. We estimated 42.48080, -76.45230 with an uncertainty of about 500 m, from 4 flyovers. The Cornell Lab describes the cam as watching its bird-feeding garden at the edge of Sapsucker Woods.

Each coloured band shows where you could stand and hear one flyover's curve from the aircraft we matched it to; the colours match the table below. One flyover allows a whole band of places. The camera is where the bands cross, inside the white contours of the combined fit.

Dashed lines are the Doppler curves predicted from each aircraft's ADS-B track at the audio estimate, with only the source frequency fitted. f0 is that fitted source frequency: the tone the aircraft emits, before the Doppler shift. Fit is the match against the traced ridge, where 1 is perfect.

At the audio estimate, all three turboprop passes are clearly heard. Each of the next-best places explains only one of them. Five quieter passes within 7 km of the estimate are not heard; they were low or idling, so this is a small unexplained gap rather than a contradiction.

The whole search area

We scored every 200 m square inside the dashed circle without knowing where the camera was. Brighter means a better fit to all the flyovers together. Only one place fits them all. The best places anywhere else fit about a third as well or less, too faintly to show at this scale.

• no visible pass came closer than 4.7 km, so the fix relies on long-range curve shapes
• sound travel used a sound-speed and wind profile measured by the aircraft themselves during the recording; with a constant sound speed the answer moves about 400 m east
• engine noise from a ground machine hides the last 70 minutes of audio

Audio: the live stream on YouTube, recorded with its HLS clock. Published location: Cornell Feeders Live: about the feeder; coordinates from OpenStreetMap.

Limits

• these are two recordings from one afternoon, so they show the method can work rather than how often it does
• the ADS-B positions come from a live feed sampled about every 17 seconds, so tracks are interpolated
• both cameras are at public institutions and their locations are published; we did not try this on cameras at private homes

Method after "This background sound leaks your exact location" by colsto. Flight data from HEXBIRD. Imagery from HEXBIRD's satellite basemap.