Demos / track-lab

From noisy sensor data to tracks an operator can trust

What a detection and tracking system is made of

Every sensor-based detection product, whether it listens, radiates or looks, runs the same seven-layer conveyor. This page is the map: click a layer to see the question it answers, the knobs it owns, the way it fails, and a small working model of it. The other pages run the layers for real.

The conveyor

Data flows left to right; each layer answers one question and hands a smaller, more abstract thing to the next. The dashed loop is the part most demos leave out: the recorded operation feeds back as data, labels and tuning.

Where the trade-offs live

Two of the layers own almost all of the precision-versus-recall budget. The detector sets how many dots appear per scan, real and false, with one threshold. The track manager decides which of those dots become something an operator sees. The product number that matters, false tracks per hour, is decided in layer 6, not layer 2, and the same is true of lead time. A detector that is tuned in isolation optimises the wrong number.

The second thing the map makes visible is where the information is. A target with a few pixels or a few decibels of signal is invisible in one scan and obvious across twenty, because noise averages out and motion does not. That is why layer 6 makes decisions on accumulated evidence, why track-before-detect skips the threshold entirely, and why every evaluation here is at the track level.

Third, ground truth is never complete. Operators mark what they saw, late, and never what they missed. An evaluation protocol that assumes clean labels rewards the wrong tracker; the imperfect-truth study shows which metrics survive that.

The pages

track-lab · Python reference in src/tracklab, experiments in experiments/, the experiment data in docs/data/. Everything on these pages is reproducible from a seed.