applets.

Radar Tracker

← All apps

About Radar Tracker

A radar does not see a ship's course and speed. Each turn of the aerial gives one plot: a range and a bearing, each a little wrong, and now and then no plot at all. The tracker inside an ARPA or a track extractor turns that string of jittery plots into a smooth position, a course and a speed. This page opens the tracker up for a plot review.

TRACK: one contact runs steady, then alters course (or slows) at a steady rate of turn of 30° a minute (or 3 knots a minute). The radar's plots are green; two trackers follow them side by side: an alpha-beta tracker (amber) and a Kalman filter (cyan). The close-up box shows the gates (where each tracker expects the next plot; a plot outside is not used, and the gate opens wider each scan nothing is used) and the Kalman filter's own 95% ellipse, which grows on missed plots and shrinks again as plots come in. The chart below gives each tracker's error against the truth. RUN replays the scans from the start; the time slider reviews any scan; NEW SEED deals new noise (the same seed always gives the same noise, so a run can be repeated exactly). Keys: Space runs or pauses, ← and → step one scan, N deals a new seed.

GAINS: the alpha-beta tracker's two numbers and what they buy. Alpha is how far the position moves towards each new plot, beta how much of the miss goes into the velocity. Big gains follow a manoeuvre quickly but pass the noise through; small gains smooth the noise but lag behind a manoeuvre. The page shows the Benedict-Bordner choice, the critically damped choice, and the gains a Kalman filter settles on, with the noise each leaves on the speed and the time each takes to catch up with a sudden change.

TWO CONTACTS: two contacts cross close to each other. Each scan the tracker has to decide which plot belongs to which track. With nearest neighbour each track takes the plot nearest its prediction, even if the other track takes the same one; with best pairing the two plots are shared out so the total miss is least. When the contacts are inside one resolution cell (here 40 m in range and a 1.2° beam), the radar gives one plot between them. Watch for the tracks swapping contacts, or both ending up on one contact. TRY 100 SEEDS counts how often each happens.

THE STORY

The alpha-beta tracker is a fixed-gain filter used since the early days of automatic radar tracking. T. R. Benedict and G. W. Bordner published their choice of gains, β = α² / (2 − α), in 1962, picked to balance smoothing the noise against following a manoeuvre. Rudolf Kalman published his filter in 1960; it works out the best gains each scan from how uncertain the track is and how noisy the plot is, and it soon went into guidance and tracking. Paul Kalata showed in 1984 that a steady Kalman filter of this kind settles on fixed alpha-beta gains set by one number, the tracking index Λ = σaT² / σm: how far the target can wander by accelerating during one scan, compared with the measurement noise.

On merchant ships, automatic radar plotting aids (ARPA) became compulsory on larger ships under SOLAS in stages from 1984. IMO's current performance standards for radar equipment, resolution MSC.192(79) of 2004, apply to radars installed from 1 July 2008, and IEC 62388 sets out how they are tested. Among the target tracking requirements: in steady-state tracking the trend of a target's motion is to be shown within 1 minute, and its predicted motion to full accuracy within 3 minutes. The chart marks 1 and 3 minutes after the track starts and after the alteration ends. The standard's accuracy figures belong to its own test scenarios and sensor errors, so the page doesn't compare its numbers with them.

HOW IT WORKS

Plots: range error is drawn with the standard deviation you pick (in metres), bearing error likewise (in degrees), then the plot is turned into x and y. A bearing error becomes a sideways error of range × bearing error, so at 6 nm a 0.5° error is about 97 m, while 25 m of range error stays 25 m: the cloud of plots is long and thin, across the line of sight. Both trackers are told this shape (the measurement covariance) for every plot. Missed plots come at random with the chance 1 − Pd; the first two plots always arrive, and both tracks start from them (position from the second, velocity from the difference).

Alpha-beta: predict the position one scan on, compare with the plot (the residual), then new position = prediction + α × residual and new velocity = old velocity + β / T × residual. Fixed small gains would take minutes to forget the rough start, so the tracker starts with the gains of a straight-line least-squares fit over the plots so far, 2(2k − 1) / (k(k + 1)) and 6 / (k(k + 1)), and drops to the set gains once those fall below them.

Kalman filter: a constant-velocity model with random acceleration of standard deviation σa, held steady over each scan. Each scan it predicts the state and its uncertainty, then weighs the plot against the prediction by their uncertainties. A large σa trusts the plots, a small one trusts the prediction.

Gates: the gate is the 99% region where the next plot should land, worked from the predicted uncertainty plus the plot's own. A plot outside is rejected; after each scan with nothing used the gate is made wider (×2, ×3 …), and after 12 such scans the track is dropped as lost.

MANOEUVRING: each tracker adds up its last five residuals, each scaled by the spread it expected. A steady target gives a sum that rarely tops 29.6 (1 time in 1,000); above that the flag goes up and stays at least 30 seconds, until the sum falls below 16. With Raise gains when MANOEUVRING ticked, the Kalman filter's σa goes up ten times and the alpha-beta tracker's α to at least 0.5 while the flag is up, so they catch up faster and pass more noise. Real trackers use many variations of this. The printout's scan log gives each plot's d² for each tracker: its residual squared, scaled by the spread expected, so a steady target averages 2 and the gate is at 9.2.

Lag and noise: the error of a tracker has two parts, the noise it lets through and the lag behind a manoeuvre. The page runs the same trackers a second time on plots with no noise, making the same gate and flag decisions, which leaves the lag alone. The printed noise figures are the RMS errors over the steady stretch from 3 minutes in to the alteration.

WHAT IT LEAVES OUT

The radar stands still (think of a shore radar, or own ship stopped), so true and relative motion are the same; a ship's tracker works in relative motion and adds its own ship's course and speed, with their own errors. The radar measures the echo's centre perfectly apart from the noise: no glint, no wander across a big ship, no clutter, no false plots. The noise is the same on every scan and independent from scan to scan, the Kalman filter is told the true noise, and a bearing error of 0.5° is a round figure, not a quoted performance. Real ARPA trackers are proprietary and differ from these two textbook filters. Not for navigation.

FURTHER WATCHING AND READING

PLOT EXTRACTOR · TRACKER · REVIEW
CLOSE-UP
CHART REPLAY
12:00 SEED 1
CONTACT AND RADAR
START RANGE AT 4 MIN RANGE σ BEARING σ Pd, CHANCE OF A PLOT SCAN
TRACKERS
ALPHA0.15 BETA RULE KALMAN σa0.05