Challenge 2 of 3 · confirmation
Lock on
A radar screen full of false blips. One of them is a real target: it keeps coming back and flies in a straight line. Click it as soon as you see it; a wrong lock costs points. Five rounds. After each, three confirmation rules of a real tracker try the same screen.
Points, the same for all four: a lock earns 1000 minus 250 for every second since the target appeared, at least 100. Each wrong lock costs 300; three wrong locks end the round. No lock within 6 seconds: 0. A click counts as a lock within 30 m of the target; a method's confirmed track is judged the same way, every scan.
Scoreboard
How each method sees it
All three run the same tracker on the same blips: a Kalman filter for each candidate, a 99% gate, and one blip per candidate each scan (the rule the live tracker uses). A candidate starts from two blips one scan apart and at most 5 m apart. They differ only in when a candidate is believed and called a target. They are told the round's target strength and clutter, as you are, and never see which blip is real.
Textbook: M-of-N, 2 of 3. A candidate is confirmed as soon as it has 2 blips in its last 3 scans. Fast, and noisy: two clutter blips that happen to line up are enough.
Popular: M-of-N, 4 of 5. A common stricter setting: 4 blips in the last 5 scans. Fewer false confirmations, a slower lock.
Ours: a sequential test. Each scan, a blip in the gate adds evidence and an empty gate takes some away, weighted by how likely a real target of this strength is to show up and how likely clutter is to fall in the gate. The candidate is confirmed when the evidence passes a threshold set for a one-in-a-million chance of confirming clutter (Wald's SPRT, track-lab's setting). It does not use the blip's brightness, which you do not see either.
Over many runs
Five rounds are five samples. The button plays 500 new rounds at each target strength in your browser, at the clutter density of track-lab's experiment E2, and scores the three rules with the same points.
Measured in track-lab
Experiment E2 ran the Python tracker with the same rules on a busier sky, with many targets and over an hour of data per setting, and counted false tracks per hour and the time from a target's appearance to its first confirmed track. Its time to confirm is the same quantity as the lock time above.
Source: experiment E2 of track-lab (public/demos/track-lab/data/e2.json: rows at snr_db 9 with their policy and params, false_tracks_per_hour, false_tracks_per_hour_se, time_to_confirm_median_s, time_to_confirm_p90_s; params tau, n_cells, sigma_xy_m, pd_per_scan, expected_clutter_per_scan, sprt_pfa_gate, match_radius_m). The tracker and the rules are the engine's (tracklab.js: makeTracker, trackerStep, makeConfirm, classifyTrack), with E2's two-blip start added around it.
Where ours still fails
It does not beat M-of-N; it is another point on the same curve. A stricter rule makes fewer false locks and locks later, and the sequential test obeys the same trade-off. At E2's clutter, false locks are rare for every rule, so speed decides the points: the textbook 2 of 3 scores highest, ahead of ours.
On a busy screen it pays with time. Told that the clutter is four times denser, the sequential test asks for more evidence and makes almost no false locks, but it becomes the slowest of the three and scores the fewest points.
On a faint target it rarely locks in time. At 6 dB the target is missing from most scans. The sequential test waits for evidence that comes slowly and misses most 6-second windows; the textbook rule, which needs only two blips, locks more often.
e2.json as the page loads.