Demos / track-lab

Two failure modes · the same data for every method

Where simple trackers break

A tracker that follows straight, well-separated targets can still fail at the two things real targets do: turn hard, and pass close to each other. Each section runs several methods on exactly the same detections, replays one recorded run side by side, and then gives the measured rates over many runs.

A sharp turn

One target flies straight, turns hard, then flies straight again. A tracker predicts where the target will be next and looks for it there, and how wide it looks is a setting: too narrow and a hard turn costs the track, too wide and every straight leg costs accuracy. One recorded run first, the same detections for every method:

0.0 s
true path first track second track: the first was lost third track

–

What each method does

    Measured over many runs

    turn rate
    signal
    methodkept one track through the turnstraight-leg errorturn errorturn, worst 5%pieces per target

    –

    –

    –
    –

    –

    Two targets crossing

    Two targets fly towards each other at a shallow angle and pass within a few metres. Near the crossing both tracks see both detections, and the tracker has to decide, scan by scan, which detection belongs to which track. Get it wrong once and the identities swap: the track that followed one target now follows the other, with nothing on the screen to say so. One recorded run first:

    0.0 s
    true paths, A and B track that started on A track that started on B any later track

    –

    What each method does

    • Greedy gives each track the closest free detection, closest pair first.
    • GNN (global nearest neighbour) solves the whole assignment at once: the best total fit for this scan.
    • JPDA does not commit near a crossing: it updates each track with a blend of the nearby detections, weighted by how likely each one is to be its own.

    One would expect the best total fit to be the safest choice at a crossing. It is not automatically so: the best fit for one scan is not the same thing as the best identity history. The measured rates below decide, not the expectation. They start at the shallowest crossing, where the methods differ most; wider angles are one click away.

    Measured over many runs

    crossing angle
    signal
    methodidentities swappedboth identities keptidentity switches per run

    –

    –

    –
    –

    –

    track-lab · the runs, the replays and every number on this page come from the Python experiments E8 and E9 (e8.json, e9.json), read by the page as it loads; nothing here is simulated in your browser.