Optimal Design of α-β-(γ) Filters

Dirk Tenne Tarunraj Singh,

Center for Multisource Information Fusion
State University of New York at Buffalo
Buffalo, NY 14260

Abstract Throughout this work closed form solutions are deter-

mined for the transient maneuver error, the steady
Optimal sets of the smoothing parameter (α, β and state error and the noise ratio of an α-β-γ filter to
γ) are derived for a sampled data target tracker. A gauge its tracking performance. These metrics are
constrained parameter optimization problem is for- exploited to optimally select the set of the smoothing
mulated for specific trajectories of the target, which parameters which minimize the noise transmission ca-
includes the noise in the measurement, the steady pability and the tracking error. The evaluation of the
state error and the transient response of the filter. maneuver error is based on certain target trajectories
This work considers two classes of target trajectories, like straight line maneuvers and targets moving on a
circular and straight line maneuvers. Closed form ex- circular path which are used to construct various cost
pressions for the steady state error and the sum of functions for the constrained optimization algorithm.
the square of the errors capturing the transient be-
havior are derived for a straight line trajectory where
the target moves with constant acceleration and a 2 Performance Metrics
circular trajectory with the target moving at con-
stant speed. The constrained optimization includes Unlike the α-β tracker, the α-β-γ filter is capable of
besides the maneuver errors a metric of noise capac- tracking an accelerating target without steady-state
ity strength, expressed by the mean square response error. Besides predicting the position, the velocity is
of the filter to white noise. An optimal selection of α, also predicted by the α-β-γ filter. Its equations are
β and γ parameters is provided for various penalties given by:
on the noise filtering.
1
xp (k + 1) = xs (k) + T vs (k) + T 2 as (k) (1)
2
vp (k + 1) = vs (k) + T as (k) , (2)
1 Introduction
where the smoothed parameters are derived with the
previous prediction and the weighted innovation as
Target tracking and predicting is realized in track- follows:
while-scan systems, which are sampled data filters,
based on previously observed positions containing xs (k) = xp (k) + α(xo (k) − xp (k)) (3)
measurement noise. The performance of these filters β
is a function of its noise smoothing behavior and its vs (k) = vp (k) + (xo (k) − xp (k)) (4)
T
transient system response. However, these are com- γ
peting metrics and a tradeoff is desired based on the as (k) = as (k − 1) + (xo (k) − xp (k)) (5)
2T 2
objective of the designer.
Applying the z-Transform to Equations (1) to (5) and
A filter developed in the mid 50’s, the α-β tracker, is x
solving for the ratio xpo leads to the transfer function
popular because of its simplicity and therefore com-
in z-domain which is
putational inexpensive requirements. This permits its
use in limited power capacity applications like pas- xp
G(z) = = (6)
sive sonobuoys. The α-β filter performance has been xo
described by Sklansky [1] and various improvements α + (−2α − β + 14 γ)z + (α + β + 14 γ)z 2
have followed ([2], [3] and [4]). Sklansky’s early work z 3 + (α + β + 14 γ − 3)z 2 + (−2α − β + 14 γ + 3)z + α − 1
delineated an optimization process with the dual ob-
jectives of minimizing the noise ratio and steady state The roots of the characteristic polynomial (CP), the
maneuver error for a target on a circular trajectory. denominator of the transfer function, are required to
lie within the unit circle to guarantee stability. Jury’s Therefore, Equation (10) reduces for the α-β filter
Stability Test [5] yields the constraints on the α, β to the noise-ratio:
and γ parameters as follows:
2α2 + αβ + 2β
ρ2 = (13)
0<α<2, 0<β , 0<γ (7) α(4 − β − 2α)
4αβ
β < 4 − 2α , γ< (8) which differs from those derived by Sklansky [1] and
2−α
Benedict and Bordner [2]. Equation (10) gauges the
This volume in the α-β-γ space defines the constraints influence of noise on the output of the α-β-γ filter by
for the subsequent optimization [6]. using the unique transformation of Equation (12).

2.1 Noise Smoothing Performance 2.2 Tracking Maneuver Errors

The noise filtering capability of the tracker is charac- Besides rejecting the noise, the maneuver error needs
terized by the noise ratio, defined as the ratio of the to be minimized. In this work, we endevour to de-
root mean square value (RMS) of the system response rive closed form expressions for the steady state er-
to the RMS value of the noisy input, which is: rors and the sum of the square of the errors which
s captures the transient behavior of the α-β-γ filter.
x2p This requires the specification of predefined target
ρ≡ (9)
x2o trajectories. The two classes of trajectories used in
this work are straight line trajectories where the tar-
Since, we require the tracker to reject measurement
get is moving with constant acceleration and circu-
noise, a small value of ρ implies an excellent filtering
lar trajectories with the target moving at constant
of noise. If the input noise is assumed to be white
speed. The maneuver error is defined as the differ-
noise, the following relationship can be derived [6]:
ence between the prediction and the observed posi-
−(4ba2 + 2ab2 + 4b2 ) − 4b(1 + c)k1 tion, which is called the innovation in the smoothing
ρ2 = (10) equations (3)-(5). Similar to the transfer function
2a(b + 2a − 4)b + a(1 + c)k2
for the input-output relationship, the maneuver error
k1 = k11 c2 + k12 c + k13
transfer function can be derived as follows:
k11 = 4a + 2a3 − ab + ba2 − 6a2
k12 = ba2 + ab2 − 6ab − 2a3 + 8a (z − 1)3 xo
eαβγ = . (14)
CP
k13 = 6a2 − 2ab2 − 4ba2 + 4a − 2b2 + 7ab
k2 = (b + 2a − 4)(c2 a − ca − c2 − 2b + bc + 1) , Equation (14) reduces to the α-β filter by setting γ
to zero.
where the transformation into the a-b-c space has
been introduced. In this space, the characteristic (z − 1)2
eαβ = xo (15)
polynomial is represented as z2 + (α + β − 2)z + (1 − α)

(z + c)(z 2 + (a + b − 2)z + 1 − a), (11) 2.2.1 Circular Trajectories: Assuming

where the second order factor has a form which is that the target traces a circular path while moving
identical to the characteristic equation of the α − β at constant speed resulting in a constant angular
filter and the third pole is real and is located at velocity, the maneuver error can be resolved into
−c. Comparing the denominator of Equation (6) components ex and ey in the x and y directions
with Equation (11), the following transformation is respectively. Substituting the circular path tra-
derived: jectory into Equation (14), we can show that the
magnitude of the steady-state maneuver error is
α = 1 + c(1 − a) constant at all sampling instants, and only the phase
1 is changing with each interval [1]. The magnitude of
β = a(1 + c) + b(1 − c) (12) the steady-state error of a target on a circular path
2
γ = 2b(1 + c). is given in Equation (16). To determine the steady
state error for an α-β filter, we can substitute xo
From Equation (12), we can infer that c equals −1 in Equation (15), or set γ to zero in Equation (16).
when γ = 0, and furthermore a and b degenerate to Both lead to the maneuver error for the α-β filter,
α and β.

4R sin3 ( ωT
2 )
e= q (16)
sin2 ( ωT 2 2 2 γ
2 )[((2 − α) cos ωT + α + β − 2) + α sin ωT ] + γ( 32 (cos ωT + 1) −
α
4 sin2 ωT )
 4R sin2 ( ωT
2 )
v 2 T 3 (α − 2) vas T 4
e= q , + + . (22)
αβ(2α + β − 4) β2
[(2 − α) cos ωT + α + β − 2]2 + α2 sin2 ωT
(17)
which is the same as that derived by Sklansky [1].
3 Design of Optimal Filters
2.2.2 Straight Line Maneuver: Assuming
Selection of the smoothing parameters within the re-
the target trajectory to be a straight line, the coordi-
gion of stability is a function of the targets trajectory,
nate system can always be placed such that one of the
the noise in the measurement, the steady state error,
axes is coincident with the trajectory and the prob-
and the transient response of the filter. To arrive at
lem thus reduces to single dimension. The observed
the optimal set of parameters, a constrained parame-
position for an accelerating target (as ) with initial
ter optimization problem is formulated. This is feasi-
velocity (v) can now be determined in the z-domain,
ble since closed form expressions for various metrics
so that the definition of the error becomes:
have been derived in Section 2. Based on these met-
1 2
2 as T z(z + 1) + vT z(z − 1) rics, closed form solutions for the optimal parameter
eαβγ = . (18) are derived for certain target trajectories.
CP
Defining a cost function which consists of two terms:
From Equation (18) the steady-state error can be de- the first, is a function of the tracker error and the
rived by using the final value theorem [5]. Due to second is a function of the noise ratio. This provides
the acceleration, the α-β filter exhibits a steady-state us with the flexibility to include steady state error,
error of: transient error etc. and permits us to weight them
as T 2
eαβ (t → ∞) = , (19) based on their importance. Thus, the cost function
β
becomes:
whereas the steady-state error of the α-β-γ filter van- f ≡ f (e, ρ, κ) (23)
ishes.
where κ is a weighting factor to penalize the contri-
The transient response can be characterized by a met- bution of one term relative to the other. The con-
ric which is defined as the sum of the square of the straints for the optimization are defined by the sta-
tracking errors as time tends to infinity. This metric bility region of the tracker defined by Equations (7)
can be defined in the discrete time domain as and (8). Different metrics have been derived to cap-
∞
X ture the tracking error for specific maneuvers such as
J= T e2 (k) . (20) straight line (Equation (21)) or circular target path
k=0 (Equation (16)). The choice of the appropriate cost
It can be calculated by simulating the response of the function depends on the goal of the optimization.
transfer function given in Equation (18) and summing
the resulting error. Instead, of using simulations, a 3.1 Circular Trajectory
closed form expression can be derived using Parse- To study the variation of the smoothing parameters
val’s Theorem [5]. Similiar to the derivations of the as a function of the relative importance of the steady
noise-ratio, the α-β-γ filter is transformed to the a- state error for a circular maneuver and the noise ratio,
b-c space, where the metric of Equation (20) can be a series of optimizations have been carried out. The
shown to be: maneuver error e in Equation (23) is obtained for a
circular target path in steady state conditions. Sub-
2(c − ca − 1)T 3 v 2 stituting Equation (17) and (13) in the cost function
Jabc =
a(2a + b − 4)(c − 1)[c(a + b + ca) − (1 − c)2 ] f , Equation (23) leads to the cost function
1 2 5
2 as T (c(a − 1) − 1)
+ . (21) J = e2 + κρ2 x2o . (24)
ab(1 + c)[c(a + b + ca) − (1 − c)2 ]
Equation (21) cannot be reduced to an α-β tracker by Assuming that the smallest turn radius of a sub-
simply substituting c with −1 because this tracker marine, for instance, with v = 3 knots speed is
exhibits a steady-state error for a target moving 45 m, leads to the angular velocity of ω = v/R =
with constant acceleration. Consequently, solving for 0.034 rad/sec, and furthermore, assume that the vari-
Equation (20) requires the subtraction of the steady- ance of the measurement noise is 200 m2 . To illus-
state value, defined in Equation (19). The transient trate the effect of changing the weighting parameter,
response metric now reduces to: a series of optimizations are carried out and the re-
sulting set of α, β and γ respectively parameters are
1 2 5
4 as T (2β + αβ + 2α2 ) plotted. Figure (1) illustrates that the optimal set of
Jαβ = parameters monotonically decrease with increasing κ
αβ 3
 α−β−γ Filter 4 0

α0 = 0 const. velocity
1 α const. accel.
β =4 β 3.5 steady state
0
γ critical damped
0.8 γ =1
0

0.6 3 0.00274
0.4

0.2 2.5
Smoothing Parameters

0 1e−05
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.0428
Weighting Factor κ 2

β
α−β Filter
1 7.94e−05
1.5
α
0.8 β 0.812
α =0 0.000633
0
1 0.0001
0.6
β =4
0 0.0505
0.4 0.00846
0.5 0.243

0.2 1.54
10
0
0 0 0.2 0.4 0.6 0.8 1 1.2
0 1 2 3 4 5 6 7 8 9 10
α
Weighting Factor κ

Figure 1: Optimal Solution of α, β and γ for tracking Figure 2: Optimal Solution of α and β for a straight
a circular path line maneuver with certain values of κ

until a final value for large penalty on the noise-ratio where Jαβ,v is the transient cost for constant velocity.
is reached. The shape of the curve does not change The optimal solution is obtained by searching for the
for circular maneuvers of different radii, since the in- parameters where the gradient of f vanishes, which
crease in the radius of the trajectory is equivalent to can be rewritten as:
increasing the weighting factor. ∂(ρ2 )
" #
∂Jαβ,v 
∂α ∂α2 1
∂Jαβ,v ∂(ρ ) =0. (27)
κx2o
3.2 Straight Line Maneuver ∂β ∂β

In practice the tracker rarely reaches steady state be- Equation (27) can be satisfied only if the determi-
cause the target path is continuously changing. In- nant of the Jacobian matrix is zero since the vector
cluding the transient response of the tracker in the [1 κx2o ]0 never vanishes. Equating the determinant
cost function leads to a more realistic metric. A to zero, we can solve for β resulting in the equation:
closed form solution is given for an α-β-γ and an α-
β tracker in Equations (21) and (22). We can now α2
β= , (28)
modify Equation (24) by including the metric which 2−α
measures the transient performance, as follows:
which matches the optimal solution proposed by
f = Jαβ(γ) + κss e2αβ(γ) (t → ∞) + κρ2 x2o , (25) Benedict and Bordner [2]. In addition to Equa-
tion (28), the optimal set of the parameters α and
where Jαβ(γ) is the transient error metric and β require satisfaction of one of Equations 27, which
e2αβ(γ) (t → ∞) the steady state error of the α-β for instance, is
tracker or the α-β-γ tracker respectively. The weight-
ing factor κss adjusts the influence for the steady 0 = [2β(4βα+4α2 −4β+β 2 )]κ+2v 2 T 3 (−α2 +4α−4+β) .
state error. The effect of the velocity and acceler- (29)
ation on the optimal set of the smoothing parameter This equation determines the location of the optimal
of an α-β tracker is shown in Figure (2). The three parameters on the curve given by Equation (28) as a
curves in the graph correspond to objecitve functions function of the weighting factor κ. In Figure (2) some
which weight the noise ratio to the tracking errors for values of the weighting factor are shown. Very small
targets moving with constant velocity, constant accel- penalty on the noise ratio optimizes the settling time
eration and steady state tracking error for a constant of the tracker, which is, of course, the shortest if the
acceleration input. two poles lie at z = 0. This leads to an infinitisimaly
quick tracker response and is given at the smoothing
The solid line in Figure (2) exhibits variation of the parameters α = β = 1. Increasing the penalty of the
optimal parameters as a function of the weighting pa- noise ratio, moves the parameter α and β towards the
rameter κ for targets with constant velocity. A closed critically damped curve.
form expression of this optimal curve can be derived
as follows. The objective function in Equation (25) The optimal curve for constant acceleration is ob-
reduces for the proposed case to: tained by setting the initial velocity and the steady
state weight to zero. Optimizing with respect to the
f = Jαβ,v + κρ2 x2o , (26) transient response of accelerating targets, results in
higher values for β as shown by the dash-dot line in 4 Conclusion
Figure (2). The equation describing this curve is ob-
tained by applying the same algorithm as for the case This paper focuses on the optimal design of α-β-γ fil-
of constant velocity maneuver. It can be shown that ters. To quantify the performance of these filters, var-
the dash-dot line in Figure (2) is described by: ious metrics are defined such as noise-ratio, steady-
state maneuver error and transient response metrics.
5 1p Closed form solutions are derived in the a-b-c space
β1,2 = 3 − α ± (36 − 60α + α2 ) (30)
2 2 which can be transformed into the α-β-γ space via
a nonlinear transformation. The maneuver errors are
and the second equation that needs to be satisfied is: defined for specific target trajectories like straight line
and circular path maneuvers. These are subsequently
0 = [4β 3 (4α2 −4β+β 2 +4βα)]κ+a2s (α2 −β)(2α−4+β)2 . used in conjunction with the noise-ratio to determine
(31) a figure of demerit. In particular, a constrained pa-
rameter optimization problem is formulated, where α,
The dashed line in Figure (2), which corresponds to
β and γ are bounded to lie within the stability vol-
a cost function which includes the steady state er-
ume. Variation of the tracker parameters for different
ror, reveals that minimizing the steady state error of
weights of the cost function are studied to provide the
the tracker by increasing the weight κss , forces β to
designer with information for the optimal selection of
maximize. The two equations describing the optimal
α, β and γ parameters. Furthermore, optimal closed
solutions for the case where the objective function has
form solutions are derived for steady-state circular
a strong penalty on the steady state error are derived
and straight line path maneuvers, where straight line
as follows:
maneuvers have the target moving at constant accel-
p eration or constant velocity.
β1,2 = 2 − 2α ± 2 (1 − 2α) (32)
0 = [(4α2 − 4β + β 2 + 4βα)]2κ (33)
Acknowledgment

Figure (2) reveals that a better noise smoothening is This research was funded by ONR/SPAWAR Sys-
obtained with smaller values of β and α, conversely, tem Center grant number 150-1284A. The authors
faster response is obtained with higher values of β. are grateful to Mr. Ed. Jahn and Dr. James Linas for
The special case of constant velocity has its fastest their input.
response at α = β = 1. Varying the weighting factor
κ can also be interpreted in changing the sensor noise References
variance since both parameters multiply the noise- [1] Jack Sklansky. Optimizing the dynamic param-
ratio. eter of a track-while-scan system. RCA Laboratories,
Similiar to the α-β filter, optimal sets of the smooth- Princton, N.J., June 1957.
ing parameter of an α-β-γ filter are derived and may [2] T. R. Benedict and G. W. Bordner. Synthesis
be display likewise in the upper plot of Figure 1. of an optimal set of radar track-while-scan smoothing
A closed form expression of the optimal set of parame- equations. In IRE Transaction on Automatic Control,
ters can be derived for a straight line target trajectory volume AC-1, July 1962.
where the target is accelerating. Constructing a ma- [3] Paul R. Kalata. α − β target tracking systems:
trix of the gradients of the terms of the cost function A survey. In American Control Conference/WM12.
with respect to a, b and c and equating the determi- ECE Department, Drexel Univeristy Philadelphia,
nant to zero and solving for the parameter b, leads to Pennsylvania, 1992.
the optimal solution in the a-b-c space:
[4] W. D. Blair. Two-stage alpha-beta-gamma esti-
mator for tracking maneuvering targets. In American
(1 + c)2 (a − 1) + a2 c
b1 = (34) Control Conference/WM12. Weapons Control Divi-
c(2 − a) sion, Naval Surface Warfare Center, Dahlgren, Vir-
(1 + c)(c + 1 − a) ginia, 1992.
b2 = (35)
c [5] Katsuhiko Ogata. Discrete-Time Control Sys-
tems. Prentice-Hall,Engelwood Cliffs, New Jersey,
It can be seen that b1 reduces to the optimal solution
University of Minnesota, 1987.
of the α-β filter (Equation (28)) if c = −1, whereas b2
vanishes for this case. The two additional conditions [6] Dirk Tenne. Synthesis of target-track estima-
to be satisfied to arrive at the optimal values of the tors. Master’s thesis, Department of Mechanical &
parameters cannot easily be simplified. Aerospace Engineering, State University of New York
at Buffalo, 1998.