The state runs some fifteen hundred cameras on its roads and broadcasts every one of them, live, at fifteen frames a second, to anyone who asks. We took six, along the river and over the canal. The boxes are ours: anything that moves against what the camera has learned to expect is given a number, a class, and a line in the ledger. Nothing on this page is stored. The camera does not know it is being watched back. It runs about ten seconds behind the street, and says so.
Method. Twelve frames a second are read off the picture. A slow running
average of them is what the camera expects. A pixel is a subject when it differs from
that average and from the frame a second before, so a car that stops at the light is
not slowly eaten by the background, and the hole it leaves when it pulls away is not
taken for a car. Subjects near each other are one subject. A subject seen for ten
frames that has actually moved is a track. A track that goes still is kept as stopped;
one that vanishes and reappears where it was heading keeps its number. Size is judged
for distance. The classes are guesses from size and speed and are frequently wrong;
the smallest, near the camera, are flagged for the Order's attention.
The line. A line is ruled across each picture and every track that crosses it is
counted, by direction, per minute. Nobody has ever counted. Now somebody has.
Source. New York State Department of Transportation traffic cameras, public
broadcast over HLS, 352 by 240 at fifteen frames a second, H.264. Played by
hls.js where the
browser cannot do it alone. Keys: 1–6 switch source; P takes the
second plate off and puts it back. Streams stop when this tab is hidden.
The plate. Whatever moves is printed flat in the spot ink, a few pixels off
register, and fades over about half a second. The street is one plate; the traffic is
the other.