Demos / track-lab

Experiments E1 to E6 · the Python reference, reproducible from a seed

What the numbers say

Six experiments, each answering one question, each written so it could have come back with the opposite answer. The two that matter most for a product are the confirmation frontier and the latency budget: what a policy costs in false tracks and in seconds, and what a tracker costs per scan.

E2 · The confirmation frontier

For each confirmation policy, the two numbers an operator feels: false tracks per hour and the median time from a target's birth to its first confirmed track. Every policy in a cell tracks the same detections, so differences are paired. Lower and further left is better; the frontier is the set of policies nothing else beats on both axes.

E6 · Latency per scan

Milliseconds per scan of the reference tracker against the number of targets, at ten clutter detections per scan. The association method dominates: JPDA's per-pair probabilities cost more than the Hungarian assignment as the scene fills, and IMM doubles the filter cost on top.

E3, E4, E5 · on their own pages

E1 · The simulator against the closed form

Before any of the above can be trusted, the simulator has to reproduce the detection model it claims: pd = Q(tau − A), pfa = Q(tau). E1 sweeps the threshold at five SNRs and compares measured to closed form; the largest deviation over the grid is at the level expected from the number of draws. If this experiment failed, nothing else on these pages would mean anything.

track-lab · every figure here is make reproduce away; the JSON behind each is in docs/data/, with a paragraph per experiment in results/README.md.