I.
INTRODUCTION
Decorrelated Unbiased
Converted Measurement
Kalman Filter
STEVEN BORDONARO, Member, IEEE
Naval Undersea Warfare Center
PETER WILLETT, Fellow, IEEE
YAAKOV BAR-SHALOM, Fellow, IEEE
University of Connecticut
Converted measurement tracking is a technique that filters in the
coordinate system where the underlying process of interest is linear
and Gaussian and requires the measurements to be nonlinearly
transformed to fit. The goal of the transformation is to allow for
tracking in the coordinate system that is most natural for describing
system dynamics. There are two potential issues that arise when
performing converted measurement tracking. The first is conversion
bias that occurs when the measurement transformation introduces a
Why they
are doing
bias in the expected value of the converted measurement. The second
the work is estimation bias that occurs because the estimate of the converted
measurement error covariance is correlated with the measurement
noise, leading to a biased Kalman gain. First, previously proposed
unbiased conversions are examined. Following this, the decorrelated
unbiased converted measurement approach is presented. Results
show that to overcome conversion bias and estimation bias, an
Main purpose unbiased measurement conversion should be employed that
of study
calculates the converted measurement error covariance using the
predicted measurement. The conversion approaches are evaluated in
tracking scenarios relevant to radar and sonar measurements.
Manuscript received September 9, 2012; revised January 18, 2013, June
14, 2013; released for publication August 28, 2013.
DOI. No. 10.1109/TAES.2014.120563.
Refereeing of this contribution was handled by L. Mihaylova.
Authors’ addresses: S. Bordonaro, Naval Undersea Warfare Center,
Torpedo Systems Department, 1176 Howell St., Newport, RI 02841,
United States, E-mail: (steven.bordonaro@navy.mil); P. Willet and
Y. Bar-Shalom, Department of Electrical and Computer Engineering,
University of Connecticut, Storrs, CT 06269, United States.
0018-9251/14/$26.00 
C 2014 IEEE
IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 50, NO. 2 APRIL 2014
Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.
The converted measurement Kalman filter (CMKF) is
commonly employed to address the problem of target
tracking when the measurements are in polar or spherical
coordinates [1]. The technique involves conversion of the
raw measurement into Cartesian coordinates prior to
tracking, allowing for the use of a linear Kalman filter.
This avoids the pitfalls of the extended Kalman filter
(EKF), which include the potential for divergence and
inconsistency between the filter-calculated estimation
error covariance and the true estimation error.
When utilizing a CMKF, two sources of bias must be
examined and, if significant, eliminated. The first source
of bias is conversion bias [1, 2, 3]. Conversion bias occurs
when the conversion process introduces a bias in the
expected value of the converted measurement. Use of a
biased converted measurement violates an assumption of
the Kalman filter, leading to degraded performance.
The second source of bias is estimation bias [4–6].
Estimation bias comes about due to the practical issue that
the calculation of the converted measurement error
covariance requires the true target position, which is
unavailable in practice. A previously proposed practical
resolution to this problem is to evaluate the covariance at
the measurement. This results in correlation between the
measurement error covariance estimate and the
measurement error itself, leading to an estimation bias
when the converted measurement is used in tracking.
One of the most studied examples of the CMKF uses
polar measurements and performs tracking in Cartesian
coordinates. If the conversion from polar to Cartesian is Hypothesis
unbiased, the performance of a CMKF is superior to that
of a mixed coordinate EKF (i.e., target motion in Cartesian
coordinates and measurements in polar coordinates) [2].
Proposed approaches for conversion include the
conventional conversion, unbiased converted measurement
(UCM), modified unbiased converted measurement
(MUCM), and the unscented transform (UT). Recently, a
decorrelated version of the UCM technique (DUCM) has
been proposed to address both conversion and estimation
bias [6]. Although DUCM was developed as an extension
to the UCM conversion, it is closely related to the
approximate best linear unbiased estimator (BLUE)
developed in [7]. Section II uses the polar-to-Cartesian
conversion to analyze and evaluate the conversion bias and
estimation bias for each conversion technique. The
DUCM technique is presented in Section III and compared
to the BLUE filter. Section IV evaluates the mean square
error (MSE) performance and consistency of a CMKF
utilizing each technique (see Table I). Section V extends
the evaluation from polar to spherical coordinates. This
article unites and extends [6, 8, 14].
II. EVALUATION OF EXISTING CONVERSION
TECHNIQUES
The conversion process required to employ a CMKF
involves transforming the raw measurement into a
1431
 TABLE I multiplicative compensation is
Methods Evaluated  UCM   
xm σα2 /2 rm cos αm
Acronym Method = e . (4)
ymUCM rm sin αm
CONV Conventional
UCM Unbiased converted measurement [3] Calculation of the true measurement error covariance of
MUCM Modified unbiased converted measurement [9, 10] the UCM requires the true range and bearing, and Limitation
UT Unscented transform [11] therefore cannot be done in practice. This practical
BLUE Best linear unbiased estimator [7] concern is at the crux of the variation in performance of
DUCM Decorrelated unbiased converted measurement [6]
the various conversion techniques. The practical
implementation of the UCM approach evaluates the
covariance at the measurements [3], namely,
converted measurement and estimating the converted
measurement error covariance. The primary measures of 1 2  
rm + σr2 1 + cos(2αm )e−2σα
2
Benchmark 11
RUCM =
parameters performance include bias, MSE, and consistency. For a 2 
conversion to be unbiased, the expected value of the 2
+ eσα − 2 rm2 cos2 αm
Research converted measurement (or the result of an estimator using
statement
the converted measurement) should equal the truth. For a 1 2  
rm + σr2 1 − cos(2αm )e−2σα
2
22
RUCM = (5)
conversion to be consistent, the estimate of the converted 2 
measurement error covariance should be statistically 2
+ eσα − 2 rm2 sin2 αm
compatible with the converted measurement errors relative
to the truth. 1 2  
rm + σr2 sin(2αm )e−2σα
2
12
RUCM =
2 
A. Review of Existing Measurement 2

Conversion Techniques + eσα − 2 rm2 cos(αm ) sin(αm ).

The four conversion processes analyzed are the 3) Modified Unbiased Converted Measurement (MUCM):
conventional conversion, UCM, MUCM and UT. It can be seen that the UCM measurement conversion (4)
Baseline 1) Conventional Conversion: The conventional is derived by conditioning on the true range and bearing,
measurement conversion from polar to Cartesian while the error covariance (5) is derived by conditioning
coordinates is on the measurements. This incompatibility was pointed
    out by previous authors along with the MUCM [9, 10]:
xm rm cos αm
= , (1)  MUCM   
ym rm sin αm xm −σα2 /2 rm cos αm
= e , (6)
with the associated estimate of the converted measurement ymMUCM rm sin αm
error covariance based on linearization [12]: 1 2  
rm + σr2 1 + cos(2αm )e−2σα
2
11
RMUCM =
11
RCONV = rm2 σα2 sin2 αm + σr2 cos2 αm 2
2
− eσα rm2 cos2 αm
22
RCONV = rm2 σα2 cos2 αm + σr2 sin2 αm . (2)
  1 2  
rm + σr2 1 − cos(2αm )e−2σα
2
12
RCONV = σr2 − rm2 σα2 sin αm cos αm
22
RMUCM = (7)
2
2
Analysis of the expected value of the conventional − eσα rm2 sin2 αm
conversion in (1) shows that the conversion has a bias in 1 2  
rm + σr2 sin(2αm )e−2σα
2

the mean of the converted measurement [12]. The bias can

12
RMUCM =
2
be found by taking the expectation of the converted range 2
− eσα rm2 cos(αm ) sin(αm ).
and bearing measurements xm and ym . If the range
measurement noise w r and bearing measurement noise wα 4) Unscented Transform (UT): The unscented transform
are uncorrelated, zero mean, and Gaussian with standard approximates the mean and covariance of the converted
deviations of σ r and σ α , respectively, the expected measurement by passing five (second-order UT) or nine
converted measurement is (fourth-order UT) sigma points through the nonlinear
    function (1) and calculating a weighted mean and
(r + wr ) cos (α + wα ) −σα2 /2 r cos α
E =e . (3) covariance [11]. For the assumption that the range and
(r + wr ) sin (α + wα ) r sin α
bearing measurement errors are uncorrelated, the sigma
Evident in (3) is that there is bias along the true bearing to points for the second-order UT are given by
the target.    
r rm rm rm rm+ rm−
New
approach
2) Unbiased Converted Measurement (UCM): Given = + −
, (8)
σ α , the bias can be calculated and eliminated. Although α α m αm αm αm αm
the original proposal for bias removal used an additive where
correction term [2], the exact compensation is √ √
multiplicative [3]. The unbiased conversion with rm+ = rm + 3σr rm− = rm − 3σr , (9)

1432 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 50, NO. 2 APRIL 2014

Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.
 TABLE II
Conversion Bias

Method Bias

r e−σα /2 − 1
2
Conv
UCM 0

r e−σα − 1
2
MUCM
 √ 
2 1
r e−σα /2
2
UT + cos 3σα −1
3 3

Fig. 2. Conversion bias for conventional, UCM, and MUCM/UT

conversion methods for 10 000 measurements with true range of 3000 m
and true bearing of 0◦ , with σ α = 10◦ and σ r = 15 m. Covariance
estimates are displayed for Mahalanobis distance of 3.

Fig. 1. Histogram of 1 000 000 Monte Carlo runs using a range of

3000 m and bearing of 0◦ to 90◦ . Range and bearing measurement noises The significance of the bias is the absolute bias divided by
used were uncorrelated Gaussian with σ r = 30 and σ α = 10◦ , the standard deviation of the range measurement noise σ r .
respectively. Mean value of Monte Carlo result and ground truth are Therefore, angle accuracy σ α , range accuracy σ r , and
plotted for comparison. In all cases, UCM was unbiased. range r all play a role in how the bias impacts estimation
performance.
+
√ −
√ The claim that the UT is a biased conversion may seem
αm = αm + 3σα αm = αm − 3σα , (10) at odds with the claims of Julier [11], but it is not. The UT
and the weights W are provides an unbiased estimate of the mean of the
conversion, not an unbiased estimate of the ground truth.
[W ] = 1
3
1
6
1
6
1
6
1
6
. (11) The UT, in essence, properly captures the bias but does not
eliminate it [13]. It is also interesting to note that the series
B. Evaluation of Conversion Bias
expansions of the MUCM and UT bias at σ α = 0 are
Conversion bias is defined as the difference between equal up to, but not including, the sixth order. In this
the expected value of the converted measurement and the light, the polar-to-Cartesian UT can be considered an
truth. Table II shows the conversion bias for each approximation to the MUCM approach [14].
technique and indicates that UCM is the only unbiased Fig. 2 provides a comparison of the expected value and
conversion. expected covariances of the conventional, UCM, and
Monte Carlo simulation verifies that UCM is unbiased MUCM/UT techniques in polar and Cartesian coordinates.
for all bearing angles. Fig. 1 shows the results of Monte The expected value of the conventional conversion is
Carlo simulation for various bearing angles. An important straddled by the expected values of the UCM and
observation is that the distribution of the converted MUCM/UT techniques.
position is very close to Gaussian along the cross-range The advantage of the UCM conversion is that it is
axis, but it is non-Gaussian along the down-range axis. For unbiased, an essential attribute in state estimation, while
this distribution, the unbiased estimate is not necessarily the MUCM conversion has a lower MSE [6]. Both
the minimum MSE estimate or the maximum likelihood conversion techniques use a multiplicative term. The
Benchmark
variable estimate. multiplicative term that results in the smallest expected
For each of the biased conversions, the bias is along squared error can be derived using a factor η as follows:
the true bearing to the target and is proportional to the
range. The absolute bias increases for long ranges and    
xmMMSE rm cos αm
poor angle accuracies. The relative bias (i.e., the bias =η . (12)
divided by the true range) is a function of angle accuracy. ymMMSE rm sin αm

BORDONARO ET AL.: DECORRELATED UNBIASED CONVERTED MEASUREMENT KALMAN FILTER 1433

Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.
 Fig. 3. Bias and MSE versus scale factor η for range of 3000 m,
σ α = 10◦ , and σ r = 30 m.

The expected squared error,

E (η (r + wr ) cos (α + wα ) − r cos α)2
(13)
+E (η (r + wr ) sin (α + wα ) − r sin α)2
can be found to be
η2 (r + σr2 ) − 2ηr 2 e−σα /2 + r 2 .
2
(14)
The minimizing η, given by
r2
e−σα /2 ,
2
Intermediate variable η= (15)
r + σr
2 2

is bounded by the MUCM scaling term. Therefore, the

MSE of the MUCM conversion is always less than that of
the UCM conversion [15]. This is confirmed by Fig. 3,
which shows that UCM is unbiased, MUCM has the
minimum MSE, and the conventional conversion is a
compromise of the two.
Fig. 4. NES for UCM and MUCM conversions based on 10 000 Monte
Carlo runs using range of 3000 m and bearing of 0◦ to 90◦ . Range and
bearing measurement noises used were uncorrelated Gaussian with
σ r = 30 m and σ α = 1◦ , 5◦ , and 10◦ . Plots include chi-square 0.99
probability bounds.

C. Evaluation of Consistency
In addition to bias and MSE performance of the
conversion, consistency is an important metric. For a
conversion to be consistent, the MSE of the converted
measurement relative to the truth must be consistent with
the estimated converted measurement error covariance. A
common measure of consistency is the normalized error
squared (NES) [12],
1   −1
N
Benchmark variable NES = z̃ R z̃i ,
N i=1 i i
where z̃i is the converted measurement error, Ri is the
converted measurement error covariance estimate for trial
i, and N is the number of trials. The NES of a consistent
estimator should be close to the state dimension n, which
in this case is 2, and (if z̃i is zero-mean Gaussian with
covariance Ri ) be chi-square distributed with nN degrees
of freedom.
Both the UCM and MUCM approaches are statistically
consistent based on their NESs, while the conventional
conversion is not. The individual components of the UCM
and MUCM converted measurements, however, are
inconsistent in certain geometries. In particular, the
covariance estimate along the true bearing is
overestimated, resulting in an NES less than 1. Fig. 4
shows the total NES and the individual components’
NESs. The overestimation of the covariance is exhibited
for small angles in the x component and near 90◦ in the y
component. The UT results, not shown, are very similar to
the MUCM results, with the fourth-order UT performance
(16) being nearly identically to that of the MUCM approach.
D. Evaluation of Estimation Bias
To allow for practical implementation, each of the
conversion methods described in Section II uses the
measurement to calculate the converted measurement
error covariance. As a result, the estimate of the
covariance becomes correlated with the measurement
noise, leading to a biased estimator [4–6]. This estimation
bias can be illustrated using a linear least squares
estimator (LLSE) for a static target. The LLSE x̂ is an
average of the converted measurements, weighted by the

1434 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 50, NO. 2 APRIL 2014

Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.
 Fig. 5. Estimation bias versus σ α for 10 000-sample LLSE using UCM
and MUCM/UT conversion methods for various ranges and σ r = 50 m Fig. 6. Total bias versus σ α for 10 000-sample LLSE using UCM and
and 0.5 m. MUCM/UT conversion methods for various ranges and σ r = 50 m and
0.5 m.

inverse of the converted measurement noise covariance. A

true bearing of 0◦ is used in the evaluation to place the
bias conveniently along the x-axis. The estimation bias best
and overall bias btot are defined as
2
best,UCM = E [x̂] − E eσα /2 rm cos αm , (17)

best,MUCM = E [x̂] − E e−σα /2 rm cos αm ,

2
(18) Fig. 7. Expected values of first 10 scans of LLSE for range of 6000 m
and α = 0, σ α = 5◦ , and σ r = 50 m.
and

btot = E [x̂] − xtrue .
In the example geometry, the bias is along the x-axis,
while in general it is along the true line of sight to the
target. The bias can have a negative impact on
performance for sonar [6] and radar [8] applications. An
implementation of a LLSE using 10,000 measurements
Limitation
was used to evaluate estimation bias. Fig. 5 evaluates the
estimation bias for two cases, one applicable to sonar with
σ r = 50 m and the second applicable to radar with σ r =
0.5 m. As the figure shows, estimation bias is a problem
for all conversion methods. It is interesting to note that
while the estimation bias for each method is nearly
identical (Fig. 5), the total bias (Fig. 6) is smaller for the
MUCM and UT techniques. This is due to the fact that the
MUCM and UT conversion bias and estimation bias are in
opposite directions and have approximately the same
magnitude. While the conversion bias and estimation bias
fortuitously nearly cancel, using a biased conversion is not
ideal for recursive estimation. For trackers using low
process noise, the MUCM conversion will eventually
converge to a solution with little bias. For higher
Limitation`
process-noise situations, where the tracker relies more on
(19) the converted measurement, the biased measurement
conversion will degrade performance.
Also disconcerting for all the methods is that a static
target will, on average, initially appear to be moving away
from the sensor as the estimator transitions from the initial
conversion bias to the total bias. Fig. 7 shows the apparent
motion for the static case using the UCM conversion.
III. DECORRELATED UNBIASED CONVERTED
MEASUREMENT
An ideal measurement conversion would be unbiased
and consistent, would provide minimum MSE estimates,
and would not suffer from estimation bias. Since no single
conversion can achieve all these goals, compromises must
be made. The primary trade-off is the MSE performance
of the MUCM and UT conversions against the unbiased
characteristic of the UCM technique. Since a fundamental
assumption in the Kalman filter is that the measurement
errors are unbiased, the UCM method is preferred, and an
adjustment to the output of the filter can be made to
achieve the MSE performance of MUCM and the UT. The
measurement error covariance must also be uncorrelated

BORDONARO ET AL.: DECORRELATED UNBIASED CONVERTED MEASUREMENT KALMAN FILTER 1435

Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.
 with the measurement error to overcome estimation bias. coordinate system, namely,
New
approach DUCM has been proposed [6] to achieve these goals. 1 2 
DUCM utilizes UCM to eliminate conversion bias and 11
RDUCM = rt + σr2 + σr2t
2  2
avoids correlation of the converted measurement error
. 1 + cos(2αt )e−2σα e−2σαt eσα
2 2
covariance estimate and the measurement noise to
preclude estimation bias. To decorrelate the measurement 1  
− rt2 + σr2t 1 + cos(2αt )e−2σαt
2
error covariance from the measurement noise, the
2
measurement error covariance can be conditioned on the
previous measurement [4] or on the predicted estimate 1 2 
22
RDUCM = rt + σr2 + σr2t
(i.e., use one-step predictions). Conditioning on the 2  2
predicted estimate leads to improved performance [16]. · 1 − cos(2αt )e−2σα e−2σαt eσα
2 2

The use of the estimate versus the previous measurement

provides for improvement in two ways. First, the estimate 1  
− rt2 + σr2t 1 − cos(2αt )e−2σαt
2

is, in general, more accurate than the measurement; and 2

second, the estimate can be propagated to the time of the 1  
new measurement, while the previous measurement lags
12
RDUCM = rt2 + σr2 + σr2t
2  2
the current target position.
· sin(2αt )e−2σα e−2σαt eσα ,
2 2
Two approaches have been proposed that evaluate the (22)
measurement error covariance at the prediction. The 1   
− rt2 + σr2t sin(2αt )e−2σαt
2
implementation of [16] evaluates at either the
2
measurement or the prediction, depending on which
estimate is viewed as more accurate. This implementation, where rt and α t are the track’s predicted estimate’s range
however, utilizes the biased MUCM conversion (6). The and bearing and σα2t and σr2t are their associated variances
BLUE filter [7] was originally developed by disregarding (see Appendix A). These quantities are approximated
the Kalman-filter framework in favor of a recursive linear using a linearization of the track’s covariance, ignoring
minimum MSE estimator. The resulting filter, however, correlation between range and bearing errors. The track’s
can be reformulated as a converted measurement Kalman predicted range and approximated range variance are as
filter [17] that uses the UCM method (4) and a converted follows:

measurement error covariance calculated with the rt = xt 2 + yt 2 , (23)
predicted estimate. Defining the predicted position as xt
and yt and its associated covariance at scan k + 1, given ⎡ ∂r ⎤
t
 
observations up to and including scan k as Pp , the ∂rt ∂rt ⎢ ∂xt ⎥
equivalent BLUE error covariance for the Kalman-filter σr2t = Pp ⎢
⎣ ∂rt
⎥,
⎦ (24)
∂xt ∂yt
framework is
∂yt
1 
1 + e−2σα Pp11 + xt2 − e−σα xt2
2 2
11
SBLUE = which simplifies to
2  
1 
1 xt
+ 1 − e−2σα Pp22 + yt2
2
σr2t = 2 xt yt Pp . (25)
2 rt yt
2 x − yt
2 2
1 2
+ σr 1 + e−2σα t2 Similarly, the predicted bearing and approximated bearing
2 xt + yt2 variance are
1  yt
1 + e−2σα Pp22 + yt2 − e−σα yt2 αt = tan−1
2 2
22
SBLUE = , (26)
2 xt
1  ⎡ ∂α ⎤
+ 1 − e−2σα Pp11 + xt2
2
t
2  
∂αt ∂αt ⎢ ∂xt ⎥
1 2 2 y − xt
2 2 σα2t = Pp ⎢
⎣
⎥, (27)
+ σr 1 + e−2σα t2 ∂xt ∂yt ∂αt ⎦
2 xt + yt2
∂yt
xt yt
= e−2σα Pp12 + σr2 e−2σα 2
2 2
12
SBLUE , which simplifies to
xt + yt2  
 1 −yt
+ e−2σα − e−σα xt yt
2 2
(20) σα2t = 4 −yt xt Pp . (28)
rt xt
2
RBLUE = eσα SBLUE − Pp . (21) The difference between the BLUE and DUCM techniques
is in the assumptions and approximations for the error
The approach proposed for the DUCM technique also statistics of the predicted estimate. The BLUE-filter
conditions the UCM error covariance on the predicted approximation is based on an assumption that the error
estimate, but the derivation is in the measurements’ statistics are Gaussian in Cartesian coordinates, while the

1436 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 50, NO. 2 APRIL 2014

Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.

Fig. 8. NES for DUCM conversions based on 10 000 Monte Carlo runs
using range of 3000 m and bearing of 0◦ to 90◦ . Range and bearing
measurement noises used were uncorrelated Gaussian with σ r = 30 m
and σ α = 1◦ , 5◦ , and 10◦ . Plots include chi-square 0.99 probability
bounds.
DUCM approach assumes that the error statistics are
Gaussian in polar coordinates. The DUCM approximation
is more valid for the initial scans of the tracker, while over
time, the BLUE approximation has more validity.
A. Evaluation of Conversion Bias and Consistency
Since the DUCM conversion is exactly the UCM
conversion, it is unbiased. For the calculation of RDUCM , a
predicted estimate xt and yt is required. To simulate this in
an evaluation of the NES, the predicted estimates xt and yt
were set to the ground truth state corrupted by normally
distributed noise with σ x and σ y at 30 m with a correlation
of 0.1. Pp was set appropriately based on this noise. Under
the condition that the underlying error distribution of the
predicted estimate is normally distributed in Cartesian
coordinates, the DUCM technique is consistent overall, as
well as consistent in the individual components of the
conversion. If, however, the underlying error distribution
of the predicted estimate is normally distributed in polar
coordinates, DUCM remains consistent overall, but the
individual components exhibit inconsistencies similar to
those in the UCM, MUCM, and UT methods. Fig. 8 shows
the NES of the DUCM conversion. The NES of the BLUE
conversion, not shown, is nearly identical to that of the
DUCM conversion.
B. Evaluation of Estimation Bias
By evaluating the converted measurement error
covariance at the predicted estimate, the covariance is no
longer correlated with the measurement noise, thus
precluding estimation bias. Evaluation of the BLUE and
DUCM techniques indicates that neither the estimation
bias nor the overall bias is significant. Figs. 9 and 10 show
the estimation and total bias for DUCM in the sonar and
radar test cases.
BORDONARO ET AL.: DECORRELATED UNBIASED CONVERTED MEASUREMENT KALMAN FILTER
Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.
Fig. 9. Estimation bias versus σ α for 10 000-sample LLSE using
DUCM conversion methods for various ranges and σ r = 50 m and 0.5 m.
Fig. 10. Total bias versus σ α for 10 000-sample LLSE using DUCM
conversion methods for various ranges and σ r = 50 m and 0.5 m.
C. Improvement of MSE Performance
While the DUCM conversion is unbiased, the MSE of
the conversion is larger than that of the MUCM
conversion, as shown in Fig. 3. That figure shows the
relationship of an unbiased estimate, with a scale factor of
2
eσα /2 , and a minimum mean square error (MMSE)
estimate, with a scale factor of e−σα /2 (i.e., different by
2
e−σα ). Based on this relationship, the following scale
2
factor (29) converts the unbiased estimate into an
approximate MMSE estimate1 :
ηDUCM = e−σαt .
2
(29)
To improve the MSE performance of DUCM, this scale
factor is applied to the output of the filter. Unlike the
MUCM approach, which supplies an MMSE and biased
converted measurement to the filter, the DUCM
application of the scaling factor only to the output allows
for the Kalman-filter assumption of unbiased
measurements to be maintained while achieving MMSE
performance.
IV. APPLICATION TO CONVERTED MEASUREMENT
KALMAN FILTER
The DUCM technique can be applied to the CMKF for
improved state estimation. The DUCM CMKF is identical
to the standard Kalman filter [12], with the use of the
1 Some researchers prefer the name federated unbiased converted
measurements (FedUCM), alluding to the union of decorrelating,
debiasing, and MMSE scaling. Subsequent plots referencing DUCM
performance include the MMSE scaling.
1437
 TABLE III
CMKF Test Cases

Parameter Test Case I Test Case II

Initial target range (μ ± σ ) 4000 ± 30 m 500 ± 10 km

Target speed σ s 10 m/s 50 m/s
Sensor σ r 20 m 0.5 m
3 ◦
Sensor σ α 5◦ 15
Tracker process noise q̃ 0.44 m2 /s3 2.18 m2 /s3

converted measurement (4) and the DUCM estimate for

the measurement error covariance using (22)–(28).
The CMKF state estimate x̂ = [ xt yt ẋt ẏt ] and
associated error covariance P can be scaled as follows for
improved MSE performance:

x̂o (k + 1|k + 1) = e−σαt x̂ (k + 1|k + 1) ,

2
(30)

Po (k + 1|k + 1) = e−2σαt P (k + 1|k + 1) .

2
(31)
The scaled estimates are for output only, and are not used
as the prior for time k + 2. Therefore, the use of this
scaling is optional and application dependent. Since the
bias is along the line of sight to the target, the scaled
output is closer to the sensor than the unscaled estimate
and should be employed when minimum MSE estimates
are desirable. If, however, an unbiased estimate is
preferred, the unscaled estimate should be used.
Regardless of the implementation chosen, the unbiased
estimate is used as the prior for the next filter iteration.
To evaluate tracking performance, two scenarios are
examined, in which the
1) true initial target range r is normally distributed,
2) true initial target bearing is uniformly distributed
(–π to π),
3) true target heading is uniformly distributed (–π to
π), and
4) true target speed χ 2DOF is distributed, scaled by σ s .
The target follows a nearly constant velocity track and
What the is estimated using a CMKF with a discretized continuous
experiment
white noise acceleration model [12]. One-point
entails
initialization of the tracker is used with an initial velocity
estimate of 0 m/s and standard deviation of σ s m/s in each
component. The parameters for each test case are listed in
Table III.
Evaluation of tracking performance indicates that
CMKFs using decorrelated techniques (i.e., DUCM and
BLUE) outperform the CMKF using the conventional,
UCM, MUCM, or UT techniques in position and velocity
MSE. As expected, the performances of the MUCM and
UT conversions are nearly identical. Fig. 11 shows the
MSE comparison of the techniques for each test case. The
posterior Cramer–Rao lower bound, as defined in [18], is
shown for reference. Both decorrelated techniques nearly
achieve the bound. The conventional performance in Test
Fig. 11. CMKF position MSE, velocity MSE, and ANEES for
conventional, UCM, MUCM, UT, DUCM, and BLUE conversion
methods from 5000 Monte Carlo runs of two target tracking scenarios.
Case II was very poor and was removed from the plot to
allow for better comparison of the other techniques.
An inspection of the DUCM and BLUE filters shows
that performance is very similar. The DUCM filter, with
the scaling of (31), exhibits improved MSE performance
for initialization. After the initial updates, performance is
nearly identical to that of the BLUE filter, with the BLUE
filter having an insignificant advantage.
To ensure credibility of the methods, the average
normalized estimation error squared (ANEES)
performance is examined. The ANEES scaled by the state
dimension n is [12]
1  T −1
N
ANEES = x̃ P x̃i , (32)
Nn i=1 i i
where x̃i is the estimation error and Pi is the error
covariance for trial i. Ideally, the ANEES of a consistent
estimator is 1. Fig. 11 shows that the decorrelated
approaches are the most consistent, with an ANEES that is
approximately 1.
V. EXTENSION TO SPHERICAL COORDINATES
The results of the previous two sections can be
extended from polar to spherical coordinates for cases in
which the measurements include elevation θ. Details on
the extension of the UCM, MUCM, UT, and BLUE
conversions to spherical can be found in [3, 10, 11, 19],

1438 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 50, NO. 2 APRIL 2014

Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.
  
respectively. The conventional conversion is · 1 − cos(2αt )e−2σαt
2

⎡ CONV ⎤ ⎡ ⎤
xm rm cos αm cos θm  
· 1 + cos(2θt )e−2σθt
2
⎢ CONV ⎥ ⎢ ⎥
⎣ ym ⎦ = ⎣ rm sin αm cos θm ⎦ , (33)
CONV rm sin θm
zm
1 2 
which has an expected value of
33
RDUCM3D = rt + σr2 + σr2t
2  2
⎡ CONV ⎤ ⎡ −σ 2 /2 −σ 2 /2 ⎤
· 1 − cos(2θt )e−2σθ e−2σθt eσθ
2 2
xm e α e θ rm cos αm cos θm
⎢ ⎥ ⎢ ⎥
E ⎣ ymCONV ⎦ = ⎣ e−σα /2 e−σθ /2 rm sin αm cos θm ⎦ . (34) 1 
2 2

CONV
− rt2 + σr2t
e−σθ /2 rm sin θm
2
zm 2 
· 1 − cos(2θt )e−2σθt
2

To apply a decorrelated and unbiased technique, the

conversion of [3] is employed:
⎡ UCM3D ⎤ ⎡ σ 2 /2 σ 2 /2 ⎤ 1 2 
xm e α e θ rm cos αm cos θm
12
RDUCM3D = rt + σr2 + σr2t
⎢ UCM3D ⎥ ⎢ σα2 /2 σθ2 /2 ⎥ 4  2
⎣ ym ⎦=⎣e e rm sin αm cos θm ⎦ . (35)
· sin(2αt )e−2σα e−2σαt eσα
2 2

UCM3D 2
zm eσθ /2 rm sin θm   2
· 1 + cos(2θt )e−2σθ e−2σθt eσθ
2 2

As in the 2-D case, the converted measurement error

covariance should be calculated at the predicted 1 
− rt2 + σr2t
measurement to avoid estimation bias. Extending the
4 
methodology in [17] to three dimensions, the BLUE filter
· sin(2αt )e−2σαt
2

for spherical measurements [19] can be put into a CMKF  

form. Defining the predicted position as xt , yt , and zt and · 1 + cos(2θt )e−2σθt
2

its associated covariance at scan k + 1 given observations

up to and including scan k as Pp , the BLUE equivalent
converted measurement error covariance for the CMKF is 1 2 
13
RDUCM3D = rt + σr2 + σr2t

2
RBLUE3D = S − Pp , (36)
· cos(αt )e−σα /2
2

  2
where S is defined in [19] and · sin(2θt )e−2σθ e−2σθt eσθ
2 2

⎡ 2 2
⎤
eσαt /2 eσθt /2 0 0 1 2 
⎢ ⎥ − rt + σr2t
=⎢ 0 ⎥
2 2
⎣ 0 eσαt /2 eσθt /2 ⎦. (37) 2
· cos(αt )e−σα /2
2
2
0 0 eσθt /2  
· sin(2θt )e−2σθt
2

The extension of the DUCM technique for spherical

measurement has the form
1 2  1 2 
11
RDUCM3D = rt + σr2 + σr2t
23
RDUCM3D = rt + σr2 + σr2t
4 2
 2
· sin(αt )e−σα /2
2

· 1 + cos(2αt )e−2σα e−2σαt eσα

2 2
  2
  2 · sin(2θt )e−2σθ e−2σθt eσθ ,
2 2
(38)
· 1 + cos(2θt )e−2σθ e−2σθt eσθ
2 2

1 2 
1  − rt + σr2t
− rt2 + σr2t 2
4  · sin(αt )e−σα /2
2

· 1 + cos(2αt )e−2σαt
2
 
· sin(2θt )e−2σθt
2

 
· 1 + cos(2θt )e−2σθt
2

where rt , α t , and θ t are the predicted estimate’s range,

1 2  bearing, and elevation. Their associated variances are
22
RDUCM3D = rt + σr2 + σr2t σr2t , σα2t , and σθ2t . As in the 2-D case, these quantities are
4  2 approximated using a linearization of tracked covariance,
· 1 − cos(2αt )e−2σα e−2σαt eσα
2 2
ignoring correlation between range, bearing, and elevation
  2 errors. The predicted range and approximated range
· 1 + cos(2θt )e−2σθ e−2σθt eσθ
2 2
variance are as follows:
1 2  
− rt + σr2t rt = xt2 + yt2 + zt2 , (39)
4

BORDONARO ET AL.: DECORRELATED UNBIASED CONVERTED MEASUREMENT KALMAN FILTER 1439

Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.
 TABLE IV
3-D Conversion Bias Multiplier

Method Bias multiplier (x and y) Bias multiplier (z)

e−σα /2 e−σθ /2
2 2 −σθ2 /2
Conv e
UCM 1 (unbiased) 1 (unbiased)
e−σα e−σθ
2
e−σθ 
2 2
MUCM 
1 −σα2 /2 −σ 2 /2  √ √  2 1 √
e−σθ /2
2
UT e e θ 1 + cos 3σα + cos 3σα + cos 3σα
3 3 3

⎤⎡
xt
1 ⎢ ⎥
σr2t = xt yt zt Pp ⎣ yt ⎦ . (40)
rt2
zt
The predicted elevation an approximated elevation
variance is
zt
θt = tan−1 , (41)
st
⎡ ⎤
−xt zt
1 ⎢ ⎥
σθ2t = −xt zt −yt zt st2 Pp ⎣ −yt zt ⎦ , (42)
st2 rt4
st2
where

st = xt2 + yt2 . (43)

The predicted bearing α t and bearing variance σα2t remain

unchanged from the 2-D case (27)–(28).

A. Evaluation of Conversion Bias

As in the 2-D case, there is a multiplicative bias that
occurs as a by-product of the conventional conversion.
The UCM technique eliminates this bias, while the
MUCM and UT conversions magnify it. The bias
multiplier is different for the z component than it is for the
x and y components. Table IV shows the bias multiplier for
each technique. Again, the UT bias is an approximation of
the MUCM bias; however, the approximation is equal up
to, but not including, the fourth order for the 3-D case.
B. Evaluation of Consistency
In the 2-D case, the consistency of the conversion,
based on the NES, was a function of the true bearing to the
target. In the extension to 3-D, consistency is a function of
the true bearing and elevation to the target. In order to
visualize the NES performance, the NES is plotted as the
radius of a surface at each bearing and elevation. A
consistent conversion would be a sphere with a radius of
the state dimension at all angles. Fig. 12 shows the
consistency plots for the UCM conversion, including the
overall consistency and the consistency of each
component. Similar to the 2-D results, the x component of
the covariance is overestimated for bearings near 0◦ and
180◦ , and the y component is overestimated for bearings
near 90◦ and 270◦ . The z component is also overestimated
Fig. 12. NES for total, x component, y component, and z component of
UCM conversion at all target bearings and elevations. Range and bearing
measurement noises used were uncorrelated Gaussian with σ r = 30 m
and σ α = 5◦ . Chi-square upper (0.99) and lower (0.01) probability
bounds are indicated on color bar as arrowheads.
for elevations near 90◦ and –90◦ . In the 2-D case, despite
the inconsistencies in the individual components, the
overall NES is consistent. This is not true for the 3-D case.
The conversion is consistent for most angles, but the
overall covariance is overestimated for large elevation
angles near ± 90◦ . This, however, is not a problem for the
many systems that deal with shallow elevation angles.
Fig. 13 shows the NES for the MUCM conversion.
MUCM exhibits inconsistencies at the same bearings and
elevations as UCM, with different characteristics.
Fig. 14 shows the NES for the DUCM conversion.
This conversion is consistent overall and in each of the
individual components at all angles, except for elevations
of ± 90◦ , under the assumption that the previous estimate
is Gaussian distributed in Cartesian coordinates. To
simulate the predicted estimate in the evaluation of the
NES, xt , yt , and zt were set to the ground truth state
corrupted by normally distributed noise with σ x , σ y , and

1440 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 50, NO. 2 APRIL 2014

Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.
 Fig. 14. NES for total, x component, y component, and z component of
Fig. 13. NES for total, x component, y component, and z component of DUCM conversion at all target bearings and elevations. Range and
MUCM conversion at all target bearings and elevations. Range and bearing measurement noises used were uncorrelated Gaussian with
bearing measurement noises used were uncorrelated Gaussian with σ r = 30 m and σ α = 5◦ . Chi-square upper (0.99) and lower (0.01)
σ r = 30 m and σ α = 5◦ . Chi-square upper (0.99) and lower (0.01) probability bounds are indicated on color bar as arrowheads.
probability bounds are indicated on color bar as arrowheads.

σ z at 30 m with a correlation of 0.1. Pp was set

appropriately based on this noise.

C. DUCM CMKF WITH SPHERICAL MEASUREMENTS

The DUCM CMKF from Section IV can be expanded
to spherical measurements with the use of the 3-D
measurement conversion (35) and associated error
covariance RDUCM3D using (38–43).
To scale the state estimate x̂ and associated covariance
for improved MSE performance (for output only), the
scaling matrix is defined as
λ1 = e−σαt e−σθt ,
2 2
(44)

λ2 = e−σθt ,
2
(45)
⎡ ⎤
λ1 0 0 0 0 0
⎢0 0⎥
⎢ λ1 0 0 0 ⎥
⎢ ⎥
⎢0 0 λ2 0 0 0⎥
=⎢
⎢0
⎥, (46)
⎢ 0 0 λ1 0 0⎥ ⎥
⎢ ⎥
⎣0 0 0 0 λ1 0⎦
0 0 0 0 0 λ2 Fig. 15. CMKF position MSE, velocity MSE, and ANEES for UCM,
MUCM, UT, DUCM, and BLUE conversion methods from 5000 Monte
and applied as follows: Carlo runs of two target tracking scenarios.

x̂o (k + 1|k + 1) = x̂ (k + 1|k + 1) , (47)

Section IV was repeated for the 3-D case with the
 following modifications:
Po (k + 1|k + 1) = P (k + 1|k + 1) , (48)
where x̂ = [ xt yt zt ẋt ẏt żt ] . The CMKF  1) true initial target elevation uniformly distributed
performance evaluation for the two scenarios in from − π3 to π3 ,

BORDONARO ET AL.: DECORRELATED UNBIASED CONVERTED MEASUREMENT KALMAN FILTER 1441

Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.
 2) true target pitch uniformly distributed from – π
to π,
3) true target speed χ 3DOF distributed, scaled by σ s ,
and
4) sensor elevation measurement with standard
deviation σ θ = σ α .
What was
Evaluation of tracking performance indicates that the
concluded CMKF using the DUCM or BLUE technique outperforms
the CMKF using UCM or MUCM in position MSE
(Fig. 15). Velocity MSE was also the best for the CMKF
using the decorrelated techniques for all but the initial
scans of Test Case I. Based on the ANEES, the
decorrelated CMKFs maintained consistency, with an
ANEES approximately equal to 1.
VI. CONCLUSION
An ideal tracker provides state estimates that are
unbiased, have minimum MSE, and are consistent. For a
converted measurement tracker to achieved idealized
results, two sources of bias need to be evaluated and
eliminated. The first is measurement conversion bias that
occurs when the conversion process introduces a bias in
the mean of the converted measurement. The second is
estimation bias that occurs when the estimate of the
measurement error covariance is correlated with the
measurement noise, leading to a biased Kalman gain. It
has been shown for converted measurement tracking
problems that a decorrelated version of unbiased converted
measurement (DUCM) exhibits improved performance
over many previously proposed techniques, and equivalent
performance to the BLUE approach. The results show that
decorrelation of the measurement error covariance from
the measurement noise is key to improved performance.
Tracker performance has been demonstrated for
estimation of position and velocity for sonar and radar
applications with polar and spherical measurements.
APPENDIX A. DEVELOPMENT OF MEASUREMENT
ERROR COVARIANCE EVALUATED AT THE
PREDICTION
11
The RDUCM term is derived here (the remaining terms
can be derived in a similar manner):
11
RDUCM = E{[rtrue cos (αtrue )
2 In IEEE Radar Conference Proceedings, ser. Proc. IEEE,
−eσα /2 rm cos (αm )]2 }.
Since the true range and bearing are not available in
practice, the estimated range rt and bearing α t are used.
Assuming the estimate errors in range and bearing wrt and
wαt , as well as the measurement errors wr and wα , are
uncorrelated, zero mean, and Gaussian with standard
deviations of σrt , σαt , σ r , and σ α , the following assertions
can be made:
rt = rtrue + wrt rm = rtrue + wr
αt = αtrue + wαt αm = αtrue + wα .
1442 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 50, NO. 2 APRIL 2014
Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.
Substituting (50) into (49),
11
RDUCM = E{[(rt − wrt )cos(αt − wαt )
2
−eσa /2 (rt − wrt + wr )cos(αt − wαt + wα )]2 }.
(51)
Taking the expectation of (51) gives the expression in (22).
ACKNOWLEDGMENT
Supported in part by grants ARO-W911NF-10-1-0369,
ONR-N00014-09-1-0613, ONR-N00014-10-1-0029, and
ONR-N00014-10-1-0412.
REFERENCES
[1] Bar-Shalom, Y., Willett, P., and Tian, X.
Tracking and Data Fusion: A Handbook of Algorithms. Storrs,
CT, USA: YBS Publishing, 2011.
[2] Lerro, D., and Bar-Shalom, Y.
Tracking with debiased consistent converted measurements
versus EKF,
IEEE Trans. On Aerospace and Electronic Systems, 29, 3
(1993), 1015–1022.
[3] Mo, L., Song, X., Zhou, Y., Sun, Z., and Bar-Shalom, Y.
Unbiased converted measurements for tracking,
IEEE Trans. On Aerospace and Electronic Systems, 34, 3
(1998) 1023–1027.
[4] Mei, W., and Bar-Shalom, Y.
Unbiased Kalman Filter using converted measurements:
Revisit,
In Proceedings of SPIE Conference on Signal and Data
Processing of Small Targets, ser. Proc. SPIE, O. E. Drummond
and R. D. Teichgraeber Eds., vol. 7445-38, San Diego, CA,
USA, 2009, pp. 1–9.
[5] Fränken, D.
Consistent unbiased linear filtering with polar measurements,
In Proceedings of 10th International Conference on
Information Fusion, ser. Proc. ISIF, vol. 1189, Quebec,
Canada, 2007, pp. 1–8.
[6] Bordonaro, S., Willett, P., and Bar-Shalom, Y.
Tracking with converted position and Doppler measurements,
In Proceedings of SPIE Conference on Signal and Data
Processing of Small Targets, ser. Proc. SPIE, O. E. Drummond
Ed., vol. 8137-12, San Diego, CA, USA, 2011, pp. 1–14.
[7] Zhao, Z., Li, X. R., Jilkov, V., and Zhu, Y.
Optimal linear unbiased filtering with polar measurements for
target tracking,
In Proceedings of the Fifth International Conference on
Information Fusion, vol. 2. Annapolis, MD, USA: IEEE,
2002, pp. 1527–1534.
[8] Bordonaro, S., Willett, P., and Bar-Shalom, Y.
Unbiased tracking with converted measurements,
(49) Atlanta, GA, USA, May 2012, pp. 0741–0745.
[9] Miller, M., and Drummond, O.
Comparison of methodologies for mitigating coordinate
transformation bias in target tracking,
In Proceedings of SPIE Conference on Signal and Data
Processing of Small Targets, vol. 4048, Orlando, FL, USA,
2000, pp. 414–427.
[10] Duan, Z., Han, C., and Li, X. R.
Comments on unbiased converted measurements for tracking,
IEEE Trans. on Aerospace and Electronic Systems, 40, 4 (Oct.
2004) 1374–1377.
[11] Julier, S., and Uhlmann, J.
A consistent, debiased method for converting between polar
(50) and cartesian coordinate systems,
 In Proceedings of AeroSense: Acquisition, Tracking and [16]
Pointing XI, Orlando, FL, USA, Oct 1997, pp. 110–121.
[12] Bar-Shalom, Y., Li, X. R., and Kirubarajan, T.
Estimation with Applications to Tracking and Navigation.
New York & Boston: John Wiley and Sons, 2001.
[13] Jorgensen T., and Rothrock, R. [17]
Removal of bias due to propagation of estimates through
nonlinear mappings,
In Proceedings of SPIE Conference on Signal and Data
Processing of Small Targets, ser. Proc. SPIE, O. E. Drummond
Ed., vol. 6969, no. 1, Orlando, FL, USA, Mar. 2008.
[14] Bordonaro, S., Willett, P., and Bar-Shalom, Y. [18]
Bias elimination in tracking with converted position and
Doppler measurements,
In IEEE Conference on Decision and Control, ser. Proc. IEEE,
Maui, HI, USA, Dec. 2012, pp. 4089–4094.
[15] Luginbuhl, T., and Hempel, C. [19]
Converting bearings-only measurements to Cartesian
coordinates,
IEEE Trans. on Aerospace and Electronic Systems, 45, 1 (Jan.
2009) 393–404.
Spitzmiller J., and Adhami, R.
Tracking with estimate-conditioned debiased 3-D converted
measurement,
In IEEE Aerospace Conference, ser. Proc. IEEE, vol. 1066,
Big Sky, MT, USA, Mar. 2010, pp. 1–16.
Fränken, D.
Some results on linear unbiased filtering with polar
measurements,
In IEEE International Conference on Multisensor Fusion and
Integration for Intelligent Systems, Heidelberg, Germany:
IEEE, 2006, pp. 297–302.
Tichavsky, P., Muravchik, C., and Nehorai, A.
Posterior Cramer-Rao bounds for discrete-time nonlinear
filtering,
IEEE Transactions on Signal Processing, 46, 5 (May 1998)
1386–1396.
Zhao, Z., Li, X. R., and Jilkov, V.
Best linear unbiased filtering with nonlinear measurements for
target tracking,
IEEE Transactions on Aerospace and Electronic Systems, 40,
4 (Oct. 2004) 1324–1336.

Steven Bordonaro received his B.S. degree in electrical engineering from Boston
University in 1991 and his M.S. degree, also in electrical engineering, from the
University of Massachusetts. He is currently pursuing a Ph.D. degree at the University
of Connecticut. He has worked at the Naval Undersea Warfare Center in Newport, RI,
since 1991. His primary areas of research for the Navy and at the University of
Connecticut have been signal processing, classification (machine learning), and
tracking.

Peter Willett received his B.A.Sc. (engineering science) degree from the University of
Toronto in 1982 and his Ph.D. degree from Princeton University in 1986. He has been a
faculty member at the University of Connecticut ever since, and since 1998 has been a
professor. He was awarded IEEE Fellow status effective 2003. His primary areas of
research have been statistical signal processing, detection, machine learning, data
fusion, and tracking. He was the editor-in-chief for IEEE Transactions on Aerospace
and Electronic Systems and continues as the vice president for publications in the
Aerospace and Electronic Systems Society (AESS).

BORDONARO ET AL.: DECORRELATED UNBIASED CONVERTED MEASUREMENT KALMAN FILTER 1443

Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.
 Yaakov Bar-Shalom was born on May 11, 1941. He received B.S. and M.S. degrees
from the Technion, Israel Institute of Technology, in 1963 and 1967, and a Ph.D. degree
from Princeton University in 1970, all in electrical engineering. From 1970 to 1976 he
was with Systems Control, Inc., in Palo Alto, CA. Currently he is the Board of Trustees
Distinguished Professor in the department of electrical and computer engineering and
the Marianne E. Klewin Professor in engineering at the University of Connecticut. He is
also the director of the ESP (Estimation and Signal Processing) Lab. His current
research interests are in estimation theory, target tracking, and data fusion. He has
published over 450 papers and book chapters in these areas and in stochastic adaptive
control. He coauthored the monograph Tracking and Data Association (Academic
Press, 1988); the graduate texts Estimation and Tracking: Principles, Techniques and
Software (Artech House, 1993) and Estimation with Applications to Tracking and
Navigation: Algorithms and Software for Information Extraction (Wiley, 2001); and the
advanced graduate texts Multitarget-Multisensor Tracking: Principles and Techniques
(YBS Publishing, 1995) and Tracking and Data Fusion (YBS Publishing, 2011). He
also edited the books Multitarget-Multisensor Tracking: Applications and Advances
(Artech House, Vol. I, 1990; Vol. II, 1992; Vol. III, 2000). He has been elected a Fellow
of IEEE for “contributions to the theory of stochastic systems and of multi-target
tracking.” He has been a consultant to numerous companies and government agencies,
and originated the series of Multitarget-Multisensor Tracking short courses offered via
UCLA Extension, at government laboratories and private companies, and overseas.
During 1976 and 1977 he served as an associate editor of IEEE Transactions on
Automatic Control, and from 1978 to 1981 he was an associate editor of Automatica. He
was the program chairman of the 1982 American Control Conference (ACC), general
chairman of the 1985 ACC, and cochairman of the 1989 IEEE International Conference
on Control and Applications. From 1983 through 1987, he served as chairman of the
Conference Activities Board of the IEEE Control Systems Society (CSS), and from
1987 through 1989 he was a member of the board of governors of the IEEE CSS. He
was a member of the board of directors of the International Society of Information
Fusion (ISIF; 1999–2004) and served as the general chairman of FUSION 2000, the
president of ISIF in 2000 and 2002, and the vice president for publications from 2004
through 2013. In 1987 he received the IEEE CSS Distinguished Member Award. Since
1995 he has been a distinguished lecturer of the IEEE Aerospace and Electronic
Systems Society (AESS) and has given numerous keynote addresses at major national
and international conferences. He was a corecipient of the M. Barry Carlton
Award for the best paper in IEEE Transactions on Aerospace and Electronic Systems in
1995 and 2000 and the recipient of the 1998 University of Connecticut AAUP
Excellence Award for Research. In 2002 he received the J. Mignona Data Fusion Award
from the Department of Defense JDL Data Fusion Group. He is a member of the
Connecticut Academy of Science and Engineering. In 2008 he was awarded the IEEE
Dennis J. Picard Medal for Radar Technologies and Applications, and in 2012 he was
awarded the Connecticut Medal of Technology. He has been listed by
academic.research.microsoft (top authors in engineering) as number 1 among the
researchers in aerospace engineering based on the citations of his work.

1444 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 50, NO. 2 APRIL 2014

Authorized licensed use limited to: University of Pretoria. Downloaded on January 22,2024 at 10:36:50 UTC from IEEE Xplore. Restrictions apply.