Challenge 1 of 3 · prediction
Where will it be
A target flies straight, then starts to turn. You see its noisy detections scan by scan, with a few false ones. The playback stops just after the turn begins. Click where the target will be two seconds later. Five rounds, turning left or right, gently or hard. After each round three filters answer on the same detections.
Scoreboard
How each method sees it
All three start from the target's first two detections and are handed its own detections for the first second. After that each one, every scan, takes the detection closest to where it expects the target, inside a 99% gate, the same rule the live tracker uses. A scan with nothing in the gate is skipped. Then each says where the target will be two seconds after the last scan.
Textbook: a Kalman filter with constant velocity. It assumes the target keeps its speed and heading, and it expects very little acceleration (process noise q = 1). It is smooth on a straight line, slow to believe a turn, and its forecast is a straight line.
Popular: the same filter with high process noise. The usual practical fix is to raise q to 30, so the filter follows a manoeuvre quickly. That buys the turn and costs smoothness on straight flight. The forecast is still a straight line, along the newest heading.
Ours: an IMM with three motion models. Three filters run side by side: straight, a left turn and a right turn, both at 0.3 rad/s. Each scan, the detection updates how likely each one is. The forecast is the average of the three models' forecasts, weighted by those probabilities.
Over many runs
Five rounds are five samples. The button runs the three filters, and the two-model IMM for reference, on 500 new scenes in your browser: 400 with a turn drawn like the rounds, and 100 with no turn at all, where the target just keeps flying straight.
Measured in track-lab
Experiment E8 ran the full Python tracker through the same manoeuvre, 200 runs per setting, and measured whether each filter kept one track through the turn and how far its estimate was from the truth. It measures the filter, not a two-second forecast, so the numbers are not the same quantity as the table above; the order of the methods is what carries over.
Source: experiment E8 of track-lab and its turn replay (public/demos/track-lab/data/e8.json, rows at snr_db 12: continuity_rate, continuity_ci_low, continuity_ci_high, turn_rmse_m, straight_rmse_m). The filters here are the engine's (tracklab.js: kfPredict, kfUpdate, immPredict, immUpdate).
Where ours still fails
Right after the onset. A few scans into the turn, nothing in the detections says "turn" yet, so the IMM still gives most weight to straight flight. Its forecast is about as far off as the Kalman filters'.
No turn at all. On a target that keeps flying straight, the IMM keeps some weight on the two turn models, and its forecast bends a little. The textbook filter, which only knows straight lines, is closer.
A turn harder than its models. The turn models are fixed at 0.3 rad/s. At 0.6 rad/s the IMM bends the forecast, but not enough.
E8 shows the other side of the fixed models: an IMM with only one turn model (straight and left) fails a right turn outright. Adding the mirrored model fixes that, and it costs something on the turn the two-model IMM was built for: a little accuracy in E8, and a clearly larger forecast error on left turns in the table of 500 runs above.
e8.json as the page loads.