Challenges

Challenge 3 of 3 · detection

Spot the drone

A live, noisy sky. At some moment a faint drone appears and flies straight across it. Click the drone as soon as you see it. Five rounds, each fainter than the last. Three detectors watch exactly the same frames, and each points at the drone once it believes it.

Points, the same for all four: a spot earns 1000 minus 200 for every second since the drone appeared (at least 100). Clicking anywhere else costs 300; three misses end the round. A second wrong point at the same spot within 1.6 s counts once, for you and for the detectors (a noise peak stays up for several frames).

the drone's path you one frame averaged TBD

Scoreboard

How each method sees it

Every frame, each detector turns the sky into a map and looks at its strongest pixel. If that pixel is above the detector's threshold, the detector points there, like a click. The thresholds are set so that on an empty sky each one points at noise about once every 100 frames (10 seconds). These thresholds are the game's own: E3 used looser ones, one false pixel per frame, because it looks only once.

Textbook: threshold the current frame. It looks at one frame at a time. A drone fainter than the noise is simply not there for it.

Popular: average the last 16 frames. Averaging lowers the noise, but it has no idea the drone moves: the drone's light is smeared along its last 1.6 seconds of flight, and the peak lands behind the drone, where it was.

Ours: track-before-detect (TBD). It adds up the last 16 frames along 49 guesses of the drone's velocity, from standing still to 0.75 pixel a frame in each direction, and keeps the best guess at every pixel. Along the right guess the drone adds up and the noise does not. The integration is the one track-lab measured in experiment E3.

Over many runs

Five rounds are five samples. The button plays 200 new rounds in your browser, 40 at each of the five levels, and records whether each detector spotted the drone and how long it took. To fit that into a few seconds, the detectors here look at a 25 x 25 pixel patch around the drone, from the moment it appears, rather than the whole sky: false alarms far from the drone are left out, so these rounds are a little easier than the game.

Measured in track-lab

The last column is experiment E3: the chance of detecting the drone in one 16-frame look, with one frame alone and with TBD. The setups differ: E3 looks once and counts a detection when any pixel of a 3 x 3 block at the drone is above a looser threshold; here the detectors run every frame, must point at the drone with their strongest pixel, and use a stricter threshold. The order of the methods is what carries over.

Source: experiment E3 of track-lab and its track-before-detect demo (public/demos/track-lab/data/e3.json, curves at L = 16: recall_frame, recall, tau_frame, tau_tbd). The sky and the integration follow the engine (tracklab.js: renderFrames, tbdIntegrate).

Where ours still fails

A drone faster than its guesses. TBD's fastest guess is 0.75 pixel a frame. A drone at 1.4 pixels a frame runs ahead of every guess, so its light never adds up in one place and the peak lands on the smear. In our check (seeds 0 to 39, the patch rule above), TBD spotted 40 of 40 drones at normal speed at 6 dB, but only 8 of 40 fast ones. At 9 dB the plain single-frame threshold spotted 24 of 40 fast drones and TBD 19, and TBD pointed at the wrong place 42 times across those 40 rounds. The fast drone flies for 4 s.

What does not break it here: a drone flying a tight circle (0.6 rad/s). In 1.6 seconds it turns only about 55 degrees, close enough to a straight line; TBD spotted 40 of 40.

track-lab · the sky and the three detectors run in your browser on the track-lab engine; the E3 numbers are read from e3.json as the page loads.