Work / track-lab

Detection and tracking2026

track-lab: from noisy detections to confirmed tracks, measured

Targets too faint for any single frame, noise that looks like targets, labels that arrive late. A reference tracker and a simulator that knows the truth, with every decision measured and runnable in the page.

PythonnumpyscipyJavaScriptCanvas
Live tracker

The problem

Every sensor product, whether it listens, radiates or looks, has the same middle: detections come in, tracks go out to an operator. That middle is where field systems fail and it is normally invisible. You cannot see a confirmation policy; you can only read about it. And the field never hands you clean labels: an operator marks a target seconds after it appeared and never marks the one they missed, so every metric computed against those labels is biased.

What I built

A simulator that knows the truth exactly, so each stage can be measured before real data exists. The signal model is one physical knob: a target of amplitude A = 10^(SNR/20) in unit noise, a threshold tau, and therefore pd = Q(tau − A) and pfa = Q(tau) in closed form. On top of it a reference tracker: batched Kalman and IMM filters, chi-square gating, global nearest neighbour on the Hungarian algorithm and a cheap JPDA, M-of-N and Wald’s sequential test for confirmation. Then the parts that make the difference in weak-signal work: track-before-detect on image frames, bearing-only localisation with the Cramér-Rao bound and an observability test, and an evaluation module written before the tracker, including a study of which metrics keep ranking trackers correctly when labels are dropped, delayed, offset or sparse.

The pages on this site run a browser port of the reference: the detection model, the Kalman filter, the two- and three-mode IMM, gating, global nearest neighbour and the greedy baseline, both confirmation policies, and the track-before-detect integrator. Tests run the JavaScript under node against the Python: the detection model and the IMM agree to one part in a million, the assignment solver returns the same total cost as scipy’s, and the default live run is pinned so an engine edit cannot silently change it. The track-before-detect story on the home page is held to experiment E3 the same way: four thousand of its runs must land inside the experiment’s 95 percent intervals.

Two things I would tell a hiring manager

The product number that matters, false tracks per hour, is decided by the track manager, not by the detector. A detector tuned in isolation optimises the wrong number. And with the amplitude in the sequential test, the detection event carries no separate information: adding log(pd/pfa) on top of the amplitude likelihood counts the threshold crossing twice and confirms a track on a single clutter hit. That bug existed in the first version of the spec; a test found it.

What the experiments measured

Eight experiments, each written so it could have come back with the opposite answer. The confirmation frontier at 9 dB: M-of-N and the amplitude-free sequential test lie on one curve, from 24 false tracks per hour at 2.2 s lead time to 134 per hour at 1.3 s; the amplitude-aided test sits above the frontier at every setting, because under a Gaussian amplitude model two strong noise cells look like a weak target. Track-before-detect at 3 dB per frame: the single-frame detector finds the target in 2 percent of trials, sixteen integrated frames in 85 percent, at matched false alarms, and the measured hypothesis-search penalty is a third of the textbook value. Under labels two seconds late, MOTA, IDF1, HOTA and time to confirm keep their ranking; false tracks per hour does not. The reference tracker runs a hundred targets in 2.4 ms per scan with global nearest neighbour and 12.7 ms with joint probabilistic association and an IMM filter, on a laptop.

Two later experiments test the improved methods against their own failure modes, 200 paired runs per cell at 12 dB. Through a turn at 0.3 rad/s, a Kalman filter tuned for straight flight loses the track in every run; raising its process noise keeps it in 99 percent at the price of accuracy (turn error 4.4 m); an IMM whose turn model matches keeps it in all runs with the lowest error, but a two-mode IMM that knows only one turn direction loses every mirrored turn, and only the three-mode design (straight, left, right) keeps both (99.5 percent). At a turn rate it does not model, 0.5 rad/s, the three-mode IMM keeps 67 percent while a high-noise Kalman filter keeps 99: a model is only as good as the motion it expects. At a 10 degree crossing of two targets, the globally optimal assignment swaps their identities in 31.5 percent of runs against 13 percent for the greedy baseline (paired test p = 2e-6), and joint probabilistic association in none; at 45 and 90 degrees the three tie.