Traffic speed from the sound of a passing car

Because the method follows the engine. A petrol or diesel engine puts out a comb of harmonics at its firing frequency, and Doppler stretches that comb exactly; an electric car does not have one, and tire noise alone is a random hiss with no persistent pattern to stretch. The same refusal appears when a vehicle is too far away for its engine harmonics to stand above the background. The alternative would be to print a number that came from fitting noise, which is worse than saying nothing.

On synthetic drive-bys with exact ground truth — built by propagating the sound through the retarded-time relation, with ground reflection, air absorption and realistic background noise — the median error is 1.7% and the worst case 4.5%. A vehicle 25 m away came out 2.5% low, one at 2 m came out 0.7% low, and at 8 dB signal-to-noise the error grew to 4.5%. A hand-held radar is typically specified at 1-2%, so this sits just outside that, without the radar.

The reading comes out low, and the tool cannot tell. The engine note of an accelerating car rises while Doppler is pulling it down, and a slow smooth rev change is mathematically indistinguishable from a slower vehicle passing further away. Three detectors were tried and measured: the fit residual did not move, a cross-check against the loudness curve flagged a perfect reading and passed a wrong one, and fitting the two halves separately gave answers 3.5% apart when the error was 13%. So the limitation is stated rather than papered over — measure vehicles that are cruising.

Because the shape of the Doppler curve contains it. How far the pitch falls tells you the speed; how sharply it falls through the moment of passing tells you how close the vehicle was. Both come out of the same fit. It is useful in practice: it separates the near lane from the far one, and a reading whose distance is obviously wrong for your spot is a reading to throw away.

That is what the 85th percentile is for — the speed that 85 of every 100 vehicles stay below, which is the figure traffic engineers use when reviewing a limit. The PDF gives it along with the median, the distribution and every individual reading. Be honest about the sample: fewer than about thirty vehicles, or all of them measured at one quiet hour, and the percentile is not yet a description of the street.

Because browsers switch on gain control and noise suppression for voice calls, and both damage this measurement: automatic gain rewrites the loudness curve used to spot a passing vehicle, and noise suppression is designed to strip steady low-frequency sound, which is precisely the engine comb being tracked. The tool asks for them off, then checks what the browser actually did and warns if the request was ignored.

Yes, and it is often better, because you can hold the phone steady, keep it out of the wind, and re-run the same recording after correcting the air temperature. Any audio or video file is scanned end to end, every pass in it is found, and each is measured separately.