Live tracker
A sensor scans a 1000 by 700 metre field ten times a second. Targets cross it, noise cells fire, and a tracker has to decide, scan by scan, what is real. Start with a recipe, then tune it yourself: every knob is a precision-versus-recall decision, and the counters show the price of each one.
What the counters mean
Detection precision and recall are the per-scan numbers: of the dots that appeared, how many were on a target, and of the targets present, how many produced a dot. They are set almost entirely by two sliders, signal-to-noise ratio and threshold, through the closed form on the right.
False tracks per hour is what an operator actually feels. A false alarm that never becomes a track costs nothing; a false track costs trust. It is the confirmation policy, not the detector, that controls this number.
Time to confirm is the other side of the same trade. A stricter policy confirms fewer clutter tracks and confirms real targets later. In the product this is lead time: seconds an operator gets before the target is close.
A track counts as false when fewer than 80 percent of the detections it absorbed came from one target. A track can change verdict while it lives: one that starts on clutter and then captures a target becomes true, one that swaps between two crossing targets becomes false.
Recorded runs
Press Record this run after a run has settled, change something, record again. The table is the experiment: same seed, different policy, compare the last three columns.
| # | SNR | tau | clutter/scan | policy | precision | recall | false tracks/h | time to confirm | seconds |
|---|
How it works, in one scan
Every 0.1 seconds the same six steps run. Each live target and each of the sensor's resolution cells produces an amplitude: a target cell draws from a Gaussian centred on its amplitude A = 10^(SNR/20), an empty cell from a Gaussian centred on zero. Anything above the threshold becomes a detection with position noise of 8 metres. That is the whole detector; it is why the closed form on the right predicts the counters exactly.
Every track predicts where it will be (constant velocity, with process noise q for what the model does not know), and accepts detections inside its gate, an ellipse at 99 percent of the innovation covariance. The assignment between tracks and detections is solved globally with the Hungarian algorithm on the negative log-likelihood, so a detection is never counted twice and an uncertain track cannot swallow everything in sight.
Assigned detections update the Kalman filter. Unassigned detections start tentative tracks. Each tentative track runs the confirmation test: M of N counts hits in a sliding window; SPRT accumulates a log-likelihood ratio, plus a·A − A²/2 for each hit when the amplitude is used, and confirms when the ratio crosses ln((1−β)/α) or drops the track when it falls below ln(β/(1−α)). Any track that misses six scans in a row is dropped.
Two things to try. Push the threshold down at low SNR and watch the clutter and the tentative tracks bloom while false tracks per hour climbs; then tighten the policy and watch time to confirm grow instead. Then switch the amplitude off in the SPRT: without it the test needs several hits and confirms slower, but it is also less fooled by a single strong noise cell.
src/tracklab; the numbers here come from the same model, run in your browser.