Current Statistical Model
Adaptive Acceleration Statistics for Agile Targets
Classical tracking filters assume a target flies at roughly constant velocity, injecting a small amount of white process noise to absorb the occasional course change. That assumption collapses against a fighter pulling 7 to 9 g or a sea-skimming missile executing a terminal weave: the filter lags, the validation gate drifts off the true return, and the track is lost. The Current Statistical Model was formulated to keep a single-model tracker locked through exactly these events. It models acceleration as a first-order Markov (exponentially correlated) process, the same correlation structure the Singer model uses, but with two decisive changes: the process mean is non-zero and adaptive, and the driving variance is computed each scan from a modified Rayleigh distribution rather than held constant.
The central idea is that at any instant the most probable future acceleration is close to the current acceleration, not close to zero. So at each update the CSM sets the mean of the acceleration random variable equal to the one-step prediction of acceleration produced by the filter itself. The acceleration variance is then written as a function of that mean and of the physically allowable extremes, the maximum positive limit amax and the maximum negative limit amin. When the estimated acceleration sits near zero, the variance is small and the filter behaves like a tight constant-velocity tracker. When the estimate climbs toward amax, the variance swells, opening the Kalman gain precisely while the maneuver is in progress and then collapsing again once the target settles.
Because the adaptation is driven by the state estimate rather than by a hard threshold, the CSM avoids the chattering and detection delay that plague maneuver-detection-and-reinitialize schemes. The penalty is modest extra arithmetic: the acceleration variance and the discrete process-noise covariance must be recomputed every scan, and the designer must supply the maneuver frequency α (the reciprocal of the time constant τ) together with the acceleration limits. These few tuning parameters give the filter enough physical realism to hold high-dynamic tracks that a fixed-Q constant-velocity filter would shed.
Discrete State and Adaptive Variance Equations
x(k) = [ p(k), v(k), a(k) ]T (position, velocity, acceleration)
Acceleration as adaptive Markov process:
ȧ(t) = −α·[ a(t) − ā(t) ] + w(t), α = 1 / τ
Modified Rayleigh acceleration variance:
σa2 = (4 − π)/π × (amax − ā)2 (ā > 0, accelerating)
σa2 = (4 − π)/π × (ā − amin)2 (ā < 0, decelerating)
Discrete process-noise factor (sampling T):
q11 ≈ (1 / 2α5) × [1 − e−2αT + 2αT + (2α3T3)/3 − 2α2T2 − 4αT·e−αT]
Where ā = current mean acceleration (set to the one-step prediction), α = maneuver frequency, τ = maneuver time constant, w(t) = white driving noise with variance 2ασa2. Example: τ = 10 s → α = 0.1 s−1; for a 9 g target amax ≈ 88 m/s².
CSM Versus Other Tracking Models
| Model | Acceleration mean | Process-noise variance | Maneuver response | Compute cost | Best application |
|---|---|---|---|---|---|
| Constant velocity (CV) | None (2-state) | Fixed, small | Poor; lags hard | Lowest | Cruise, civil ATC |
| Singer | Zero, fixed | Fixed, uniform+impulse | Moderate | Low | Mild maneuvers |
| Current Statistical (CSM) | Adaptive (= current est.) | Adaptive, modified Rayleigh | Strong, low lag | Moderate | Agile single-model fire control |
| Interacting Multiple Model (IMM) | Per-model | Per-model bank | Excellent | High (N filters) | Mixed CV/turn/ballistic |
| Constant turn (CT) | Implicit (ω state) | Tuned to turn rate | Good for coordinated turns | Moderate | Banked-turn aircraft |
Frequently Asked Questions
How does the Current Statistical Model differ from the Singer model?
Both use a first-order Markov (exponential autocorrelation) acceleration process, but the Singer model holds a zero mean and a fixed uniform-plus-impulse density. The CSM makes the acceleration mean non-zero and adaptive, setting it equal to the current one-step predicted acceleration, and recomputes the variance each scan from a modified Rayleigh density bounded by amax and amin. Centering the statistics on the actual maneuver gives faster response and smaller position lag through a sustained high-g turn, at the cost of updating σa2 every scan and specifying the acceleration limits.
How are the maneuver time constant and maximum acceleration chosen for a CSM tracker?
The time constant τ (= 1/α) sets how long acceleration stays correlated: roughly 20 to 60 s for a slow turn, 5 to 20 s for an agile fighter, and 1 to 2 s for an evasive target, so α spans about 0.02 to 1.0 s−1. The limits amax/amin bound the modified Rayleigh density; a 9 g fighter needs amax ≈ 88 m/s², while a civil target may need only 10 to 30 m/s². Too small an α lags during a maneuver; too large makes straight-line tracks noisy. Adaptive schemes tune α from the normalized innovation.
Why does the CSM use a modified Rayleigh distribution for acceleration?
Real targets cannot accelerate symmetrically without bound, and they tend to sit near the current commanded value rather than near zero. The modified Rayleigh density encodes this by making σa2 = (4−π)/π × (amax − ā)2 during positive maneuvers, so the variance shrinks while coasting and grows as the estimate approaches amax. That adaptive variance is what lets the Kalman filter open its gain during a turn and tighten it in steady flight without an explicit maneuver-detection switch.