Doppler Ambiguity Resolution
Doppler Ambiguity Resolution
INTRODUCTION
738 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 33, NO. 3 JULY 1997
Fig. 1. Description of the multiple PRF waveform.
estimation error and the PRF parameters on ambiguity N(k) and equals 2ns Nb . Fig. 1 sketches the waveform
order measurement is particularly investigated. and summarizes the notations used in this work.
The proposed method is also compared with a
standard method. Section IV is devoted to simulations EXAMPLE 1 Let Tr = 1, Nb = 1, N(1) = 20, and
evaluating the method performance in a near-real consequently, N(2) = ¡20. The values of the PRF on
environment. Problems due to clutter reduction in this bursts 1 and 2 are
multiple PRF context are also tackled. Finally, Section 1 20
V summarizes and draws conclusions on the method. Fr (1) = = = 0:9524
Tr (1) 20 + 1
1 ¡20
II. DOPPLER FREQUENCY ESTIMATION Fr (2) = = = 1:0526:
Tr (2) ¡20 + 1
ALGORITHM
740 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 33, NO. 3 JULY 1997
f̂2p+2,r , f̂r (p) is defined as A comparison can be made with the criterion
obtained in the case where initial phases of the signal
1 at the beginning of each burst are independent. In this
f̂r (p) = angle(exp(j2¼ f̂2p+1,r ) + exp(j2¼ f̂2p+2,r ))
2¼ case, it can be shown that the maximum likelihood
(17) approach leads to the maximization of the following
function:
where angle(z) returns the phase angle in radians in ¯
2Nb ¯ ns ¡1
def X¯ X k
the range 0 to 2¼ of the complex number z. Li (f, n) = ¯ yq exp(¡j2¼fTr (k)q)
The estimation of the folded frequency fr can ¯
k=1 q=0
be improved by averaging the frequency estimates μ ¶¯¯2
f̂r (p) over p, for every couple of bursts. Here again, n ¯
£ exp ¡j2¼ q ¯ : (22)
a “circular mean” relying on the periodicity of the N(k) ¯
exponential function is computed:
The global maximizer of this function, which is equal
N
Xb ¡1 to the sum of the signal periodogram on each burst,
1
f̂r = angle exp(j2¼ f̂r (p)): (18) appears to be also an estimator of fD when initial
2¼ phases are dependent. Meanwhile, the estimation
p=0
obtained by this criterion has a larger variance than
the one given by L(f, n) which asymptotically reaches
C. Ambiguity Order Resolution the Cramer-Rao bound [6]. Better performance of
L(f, n) over Li (f, n) can also be justified by the fact
Under a white Gaussian noise assumption, the
that, for each burst taken separately, the localization
extension from the classical case of uniformly
of the maximum of the periodogram exhibits a high
sampled signals of the maximum likelihood
variance, associated to the Cramer-Rao bound of a
estimation criterion of fD , [6, ch. 13, pp. 408—413],
signal with ns points. For a reduced number of bursts,
is straightforward. It can be shown that this estimator,
the summation over k (burst index) does not allow a
or equivalently the estimator of the pair (fr , nr ), is
sufficient reduction of the variance.
the global maximizer of the function defined as the
Consider now the term for k = 2p + 1 of the sum
squared modulus of the scalar product of a vector
in (19), for a specified value of f, say f0 :
of measured samples ynk and a model of this vector
corresponding to a noiseless signal xnk with a complex s ¡1
nX
initial amplitude equal to 1. This function defined on yqk exp(¡j2¼f0 Tr (2p + 1)q)
[0, 1[£[¡nmax , nmax ], ([a, b[ is a half-closed, half-open q=0
μ ¶
interval and [a, b] is a closed interval) and denoted n
L(f, n) where f and n represent, respectively, the £ exp ¡j2¼ q : (23)
N(2p + 1)
decimal and integer part of the frequency variable,
can be written according to (12) and separating the f According to the inequalities stated in (4) and
and n term as (16), the inequality jnj < N(2p + 1) always holds.
¯ 2N Consequently, the expression (23) is equal to the
¯ b s ¡1
nX
def¯ X ¤ value of the discrete Fourier transform (DFT) of ynk
L(f, n) = ¯ ®k (f, n) yqk
¯ modulated by exp(¡j2¼f0 Tr (2p + 1)q) at the frequency
k=1 q=0
n=N(2p + 1). If we now assume that the number of
μ ¶¯¯2 signal samples is smaller than N(2p + 1), i.e.,
n ¯
£ exp(¡j2¼fTr (k)q) exp ¡j2¼ q ¯
N(k) ¯ ns · N(1) (24)
(19) the expression (23) represents the nth DFT bin of
the same signal zero-padded to N(2p + 1) points.
where ®k (f, n) is, from (8) and (9):
Similarly, the term for k = 2p + 2 of the sum in (19)
®2p+1 (f, n) = exp(j2¼fns 2p) (20) equals the inverse DFT of the same signal. With a
μ μ ¶¶ FFT algorithm, the number of complex multiplications
1 and additions to calculate the 2nmax points of this
®2p+2 (f, n) = exp j2¼fns 2p + 1 +
N(2p + 1) expression becomes hence N(2p + 1) log2 N(2p + 1).
μ ¶ Values of L(f0 , n) are obtained as the sum over k
n
£ exp j2¼ns : (21) (burst index) of the expression (23) windowed by
N(2p + 1)
the constant ®2p+1 (f0 , n) for odd terms and by the
Note that the coefficients ®k (f, n) simply correspond complex sine wave ®2p+2 (f0 , n) for even terms. As a
to the coefficients ak (f, n) for a signal with an initial consequence, the number of operations required to
complex amplitude equal to 1 and are obtained by compute L(f0 , n) on [¡nmax , nmax ] is bounded above by
replacing A by 1 in (8) and (9). 2Nb N(2Nb ¡ 1) log2 N(2Nb ¡ 1).
def sin ¼°ns Fig. 4 sketches the shape of this surface and
W(°) = exp(j¼°(ns ¡ 1)) : (27)
sin ¼° Fig. 5 represents a contour plot of L(f, n) obtained
for Nb = 2, N(1) = 20, N(3) = 80 (hence N(2) = ¡20
Consequently, the surface L(f, n) is a weighted
and N(4) = ¡80), ns = 12 and fD = 2:3. These results
sum of Dirichlet kernels centered in the plane (f, n)
suggest the following remarks.
on lines defined by
1 1) The tradeoff between the variance of f̂r and n̂r
f=¡ n and appears clearly in Fig. 4. The larger jf̂r j is, the farther
1 + N(2p + 1)
(28) from zero the maxima of L(f̂r , n) are and then the
1
f=¡ n: higher the “risk” to have a large error on nr is.
1 ¡ N(2p + 1)
2) As the N(k) tend toward infinity, the slopes of
The difficulty for the development of L(f, n) in lines defined by (28) tend toward 0 bringing kernel
(26) comes from the initial phases terms which lead to centers together and preventing ambiguity order
“interferences” between kernels. For simplification resolution. This result is immediate from Definitions
purposes, we assume that ns is large enough to (1) and (3), since it can be seen that as the N(k)
consider that the width of kernel mainlobe is small increase, the waveform becomes similar to a single
compared with the distance between kernels centers. PRF. Thus, it clearly appears that the integers N(k)
742 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 33, NO. 3 JULY 1997
have to be as small as possible according to the
constraints stated in (16) and (24).
3) This interpretation also gives a motivation for
multiple PRF use for estimation purposes. In fact,
for the same variance of f̂r and for a constant total
number of points (2Nb ns constant), the additional use
of bursts with smaller N(k) has a “delocalization” Fig. 6. Definition of fr+ and fr¡ .
effect on L(f, n) maxima, reducing the estimation
error risk on nr . f̂r in the mainlobe of L(f̂r , 0). It is important to note
that this hypothesis leads to a new upperbound for
C. False Ambiguity Order Estimation Probability † ‡
Pfaoe (n), Pfaoe (n), that also tends to Pfaoe (n) when ns
tends to infinity.
The simple approach proposed in the previous
We propose now to evaluate this last quantity. For
section illustrated, without quantifying it, the link
between the estimation error in fr and n̂r . This section that, we assume small variations of f̂r and replace
is dedicated to a precise evaluation of the probability Ls (f̂r , n) and L(f̂r , 0)1=2 by their second-order series:
of performing a false ambiguity order estimation as
Ls (f̂r , n) = a0 (n) + a1 (n)f̂r + a2 (n)f̂r2 + o(f̂r3 ) (33)
a function of the variance of f̂r . As in the previous
section, we assume a noiseless signal with a null L(f̂r , 0)1=2 = b0 + b2 f̂r2 + o(f̂r3 ): (34)
Doppler. Moreover, f̂r is supposed to be zero-mean
and Gaussian with variance ¾f2 . We denote the second-order approximations of
We define the probability of false ambiguity order Ls (f̂r , n) and L(f̂r , 0)1=2 by L̂s (f̂r , n) and L̂(f̂r , 0)1=2 .
estimation at order n for n > 0, denoted Pfaoe (n), as A brief development of the coefficients ak (n)
and bk is given in Appendix A. The validity of
the probability that L(f̂r , n), or L(f̂r , ¡n), is greater
the approximation of L(f̂r , 0)1=2 and Ls (f̂r , 1) by
than L(f̂r , 0). Note that this quantity does not quantify
the probability to estimate n, or ¡n, instead of 0. L̂(f̂r , 0)1=2 and L̂s (f̂r , 1) (respectively) around f̂r = 0
Nevertheless the expression of this last probability, has been numerically checked for different parameters
as for example a function of Pfaoe (n), is very complex configurations.
because of the dependence of the random variables If fr¡ and fr+ are the frequencies corresponding
to the intersections of the parabolas defined by the
L(f̂r , k), k = ¡nmax , : : : , nmax .
Owing to the formula for L(f, n), the computation previous polynomials L̂s (f̂r , n) and L̂(f̂r , 0)1=2 (see
‡
of Pfaoe (n) requires the solution of a nonlinear Fig. 6), we adopt the following estimate of Pfaoe (n):
system of equations in order to determine the various
intervals where the inequalities define the probability
‡
Pfaoe (n) = Pr(jf̂r j > fr¤ ) (35)
hold. The first step adopted to solve this problem is Z fr¤ Ã !
2 ¡f 2
to replace Pfaoe (n) by an upperbound with a simpler = 1¡ p exp df (36)
expression. ¾f 2¼ 0 2¾f2
For this purpose, consider the inequality: Ã !
fr¤
= 1 ¡ erf p (37)
L(f̂r , n) · Ls (f̂r , n)2 (31) ¾f 2
where Ls (f, n) is with fr¤ = min(jfr¡ j, jfr+pj) and
R x erf(x) is the error
2Nb
X function defined as 2= ¼ 0 exp(¡t2 ) dt.
def
Ls (f, n) = jW(fTr (k) + n=N(k))j: (32) Appealing to the same arguments as above, we
k=1 suppose now that the error on the ambiguity order
† estimation can be, at most, equal to one. Therefore,
Let us define now a new probability, denoted Pfaoe (n), we take in what follows Pfaoe = Pfaoe ‡
(1). It is clear that
as the probability that Ls (f̂r , n), or Ls (f̂r , ¡n), is a minimization of this quantity will correspond to a
†
greater than L(f̂r , 0)1=2 . According to (31), Pfaoe (n) minimization of the probability at other orders (n > 1).
‡
overestimates Pfaoe (n). However, as seen in the Finally, Figs. 7 and 8 represent Pfaoe (1) as a
† 2
previous section, the limit of Pfaoe (n) as ns tends to function of 1=¾f in dB for different configurations
infinity equals Pfaoe (n). corresponding to the same observation time of
The error on the folded frequency estimate being 2Nb ns = 48 points. They validate the remarks
assumed small and Gaussian, and considering the formulated from the reduced frequency/ambiguity
insight given by the previous section, see Fig. 4, order plane analysis.
we assume that the condition Ls (f̂r , n), or Ls (f̂r , ¡n), 1) For Nb = 1, the probability of false ambiguity
greater than L(f̂r , 0)1=2 is satisfied only for values of order estimation decreases with N(1), Fig. 7.
2) For a constant number of samples, addition of xnk . Figs. 9 and 10 have been obtained using 1000
blocks with smaller N(k) does not worsen the results, realizations of f̂r for each value of ¾f2 . They perfectly
see Fig. 8, even though the number of samples per suit the previous conclusions. Moreover, the values
burst is smaller. of the Pfaoe from numerical simulations are less than
those previously obtained; thus the threshold lies
Moreover, these results exhibit a threshold on near 45 dB. This confirms the fact that the theoretical
‡
1=¾f2 under which the probability of false ambiguity values of the Pfaoe (above denoted Pfaoe ) correspond to
order estimation is almost zero. The existence of an upperbound of the “true” Pfaoe .
this threshold, situated near 50 dB, is fundamental,
since this value is reachable by the classical modern IV. SIMULATIONS
frequency estimators (e.g. [7]).
To conclude this section, the results of Figs. 7 and A. White Noise Only Environment
8 have been confirmed by numerical simulations. For
The proposed method, referred to as QMLA
this purpose, a signal xnk has been generated, f̂r has (quasi-maximum likelihood algorithm), has been first
been drawn according to the Gaussian probability evaluated in a white noise only environment, this
density function with mean zero and variance ¾f2 situation corresponding to ideal clutter whitening.
and nr has been estimated using L(f̂r , n) with ynk = Mean square error (MSE) curves in f̂D and f̂r as a
744 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 33, NO. 3 JULY 1997
function of the signal-to-noise ratio (SNR) are given
to show the performances of QMLA. The SNR is
defined as
2
def A
SNR = 20 (38)
¾
where A0 is the modulus of the amplitude A (as
defined above (7)) of the complex sinusoid at the
maximum of the beam pattern of the antenna in the
case of a Swerling 0 or nonfluctuating target; A20 is
the average power in the case of a Swerling I target.
In this latter case, the probability density function of
the target echo amplitude A is
μ ¶
A A2
p(A) = 2 exp ¡ 2 , A¸0 (39)
A0 2A0
Fig. 11. Comparison of reciprocal MSE in f̂r and f̂D as function
and ¾ 2 is the white Gaussian noise variance of the of SNR for different values of N(1); Nb = 1; ns = 24; Gaussian
term bnk of (5). antenna beam pattern; 1000 signal realizations. Solid line: MSE in
In some cases the signal is Gaussian modulated in f̂D for N(1) = 100; dashed line: MSE in f̂D for N(1) = 20; dotted
order to simulate the antenna beam pattern, according and dashed-dotted line: MSE in f̂r for N(1) = 100 and N(1) = 20
to the antenna time function, such that the 2ns Nb (curves superimposed).
points of interest lie in the ¡3 dB beamwidth.
The MSE in f̂D (resp. f̂r ) is estimated by 2) Above the threshold, the estimation variance
on fr becomes sufficiently small and the Pfaoe is zero.
M
1 X k The estimation variance on fD is then only due to the
MSE = (f̂D ¡ fDk )2 (40) error on fr and diminishes linearly (in dB) as the SNR
M
k=1
increases. Any improvement in this zone can only be
where M is the number of signal realizations and achieved through a better estimation of fr .
f̂Dk is the estimate of the Doppler frequency fDk for 3) A transition zone lies between ¡5 and +10 dB.
the kth signal realization. In each case, the targets Note that the threshold observed on MSE curves
are uniformly distributed in velocity within the corresponds to the threshold seen on curves of
maximum frequency domain allowed by N(1); see Pfaoe as a function of 1=¾f2 . As soon as the SNR is
(16). The folded frequency fk is estimated on each sufficiently high to allow a good estimation of fr , the
burst with the “min-norm” algorithm [7], with a ambiguity order nr and fD are correctly estimated. On
predictor length equal to 4. MSE curves in f̂D and Fig. 11, for N(1) = 20, the threshold is 5 dB which
corresponds to (1=¾f2 )dB ' 50 dB, i.e., the value of
f̂r are given in Figs. 11, 13, and 14 for different
the threshold found in the previous section. As seen
parameters configurations and type of targets. For
above, this threshold depends obviously on the choice
more insight on the ambiguity order estimation, curves
of N(k).
of Pfaoe as a function of the SNR are also given in Fig. 11 reveals the influence of N(1) on the
Figs. 12 and 16. Note that both kinds of curves (MSE performance of QMLA. Changing N(1) from 20 to
and Pfaoe ) are necessary to gauge the performance of 100, shifts almost 5 dB the transition zone. The MSE
QMLA. The combined analysis of the MSE in f̂r and curve for the folded frequency being identical in both
Pfaoe (corresponding to each step of the estimation configurations, this discrepancy in performance is
procedure) gives a better understanding of the method perfectly illustrated by Fig. 12. The improvment is
behavior than for example the normalized MSE alone, due to false ambiguity order measurements that are
which turns out to be a poor performance indicator less frequent with N(1) = 20 than with N(1) = 100.
in our case, since both types of possible errors (in fr Above the greater of the two thresholds, Pfaoe equals
and in nr ) are mixed together. The absolute MSE in zero, and the method behaves similarly for both cases.
The influence of an increasing block number is
f̂D , combined with the MSE in f̂r or the Pfaoe , is also
investigated in the Fig. 13. In this case, the MSE
interesting and useful to gauge the performance of the curve for the folded frequency being different,
QMLA. Pfaoe is not sufficient to explain the results. The
As follows, all MSE curves exhibit a threshold transition zone is here divided in two parts. From
shape. ¡4 dB to 4 dB, the results are better for the two
1) Below the threshold (in term of SNR), both blocks case. Indeed, this zone corresponds to a
folded and unambiguous frequencies are badly MSE on the folded frequency below 20 dB where
estimated, and the error on the ambiguity order Pfaoe curves are superimposed, Fig. 7. Then the
remains high. estimation of fD depends only on the variance of f̂r .
746 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 33, NO. 3 JULY 1997
Fig. 16. Pfaoe as function of SNR for fluctuating (Swerling I)
target. 3500 signal realizations; Nb = 1; N(1) = 20; ns = 24.
Fig. 17. Reciprocal MSE in f̂D and f̂r as function of SNR.
Performance comparison of QMLA (solid line: f̂D ; dashed line:
f̂r ) and UFM (dashed-dotted line: f̂D ; dotted line: f̂r ). 1000 signal
candidates for every PRF. This clustering algorithm realizations; Nb = 3; N(1) = 40; N(3) = 60; N(5) = 70; ns = 5.
does not require, like QMLA, an a priori fixed
correlation interval. When the number of samples
per burst ns is large, and with only two PRF, both
methods seem to behave similarly. However, for a
constant time on target, as the number ns decreases,
and the number of bursts Nb increases, the results of
QMLA become better than UFM ones. The results of
the comparison (Pfaoe and MSE) for Nb = 3, N(1) =
40, N(3) = 60, N(5) = 70 and ns = 5 (i.e., a time on
target equal to 30 samples) are given in Figs. 17 and
18. The MSE in f̂D and f̂r are significantly better
for QMLA than for UFM in the range of moderate
SNR (between 5 and 15 dB): for a SNR of 10 dB, the
difference in terms of reciprocal MSE on f̂D between
the two methods is almost equal to 30 dB. These
results are explained by the Pfaoe curve in Fig. 18. For
QMLA the Pfaoe becomes equal to zero for a SNR
Fig. 18. Pfaoe as function of SNR for UFM (dashed line) and
greater than 8 dB, whereas UFM exhibits a non-zero QMLA (solid line). Nb = 3; N(1) = 40; N(3) = 60; N(5) = 70;
Pfaoe until a SNR threshold equals 18 dB. Thus, a ns = 5.
significative number a false correlations still occur
in this zone (i.e., between 8 and 18 dB), which lead As a conclusion, the performance of QMLA turns
to a high MSE in f̂D and f̂r . This false correlation out to be appreciable. The threshold in SNR required
problem, mentioned in several papers [3, 9], is mainly to get a good unambiguous estimate (reciprocal of the
due to the poor precision of the folded frequency MSE greater than 40 dB) is similar to the minimum
estimates on bursts of ns samples (when ns is small) SNR required to achieve good results in the target
which prevents an accurate correlation procedure. detection process [4]. Moreover, QMLA exhibits
This phenomenon disappears when ns is taken large superior performance over a standard technique
enough. UFM splits the signal into bursts and treats especially when the number of samples per burst is
them independently, whereas QMLA takes advantage small.
of the global signal coherence first by averaging
the folded frequency estimates on every burst to B. Clutter Environment
compute f̂r and then by evaluating the function L(f, n)
from the whole signal to estimate nr . This property Thus far, only the white noise-only environment
of QMLA seems to be its main advantage over the has been considered. Ground, rain, or sea clutter
classical techniques, especially in difficult estimation composed of powerful unwanted echoes, may however
contexts (i.e., small number of signal samples per alter notably the estimation performance. It is then
burst, moderate SNR). necessary to filter out the clutter. After clutter
748 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 33, NO. 3 JULY 1997
higher than 10 dB, whichever the SNR is (20 dB or This leads to the following expression of
100 dB). coefficients a0 (n), a1 (n), and a2 (n):
2Nb
X sin(¼nns =N(k))
V. SUMMARY a0 (n) =
sin(¼n=N(k))
k=1
A new method for non-ambiguous Doppler
2Nb
X
frequency estimation for coherent pulsed radar has sin(¼nns =N(k))
a1 (n) = r (k, n)
been presented and its performance in white-noise sin(¼n=N(k)) 1
k=1
only and noise-plus-clutter environment has been
evaluated and illustrated through both theoretical 2Nb
X sin(¼nns =N(k))
developments and computer simulations. This a2 (n) = r (k, n):
sin(¼n=N(k)) 2
new estimation scheme, based on a multiple PRF k=1
waveform, with particular repetition frequencies
values, operates directly on the entire signal and Second-Order Series of L(f̂r , 0)1=2 .
allows a fast numerical implementation. Its ability to From (19), L(f̂r , 0)1=2 can be written
use any number of different PRF (for a given signal ¯ 2N ¯
¯Xb sin ¼ f̂r Tr (k)ns ¯¯
length) lowers the probability of false ambiguity ¯
L(f̂r , 0) = ¯ exp(¡j2¼f̂r q(k)) ¯
order measurement for a given variance in the folded ¯ sin ¼ f̂ T (k) ¯
k=1 r r
frequency estimate. Nevertheless, a compromise has
to be found between the number and the length of with
blocks in order to ensure a correct global behavior q(2p + 1) = 2pns + Tr (2p + 1)(ns ¡ 1)=2
(on both the folded frequency estimation and the μ ¶
ambiguity order measurement). Its main advantage 1
q(2p + 2) = 2p + 1 + ns
over alternative techniques comes from its ability N(2p + 1)
to process directly the whole signal, hence taking
+ Tr (2p + 2)(ns ¡ 1)=2:
advantage of the signal coherence between bursts,
whereas other classical estimation methods tend The development of this expression gives
to split the signal into bursts before searching for
correlation between the unfolded frequency estimates. L(f̂r , 0) = ns j2Nb + js1 f̂r + s2 f̂r2 + o(f̂r3 )j
Because of this property, simulations have shown that
with
the paper method overcomes the false correlation (and
hence the false ambiguity order estimation) problem 2Nb
X
encountered with classical techniques and provides a s1 = ¡2¼ q(k)
significant improvement in performance. k=1
à 2N 2Nb
!
2 1 ¡ n2s Xb 2
X
2
s2 = ¼ Tr (k) ¡ 2 q(k) :
APPENDIX 6
k=1 k=1
This Appendix presents the second-order series Finally, we obtain following expressions for b0
expansion of the functions Ls (f̂r , n) and L(f̂r , 0)1=2 , and b2 :
defined in (33) and (34).
b0 = 2ns Nb
Second-Order Series of Ls (f̂r , n).
μ ¶
First, the second order series of jW(f̂r Tr (k) + s12
b2 = ns s2 + :
n=N(k))j is derived 4Nb
¡ cot(¼n=N(k))) [1] Ludloff, A., Füchter, N., Hagerdorn, F., Minker, M., and
Rohling, H. (1981)
r2 (k, n) = ¡r1 (k, n)¼Tr (k)cot(¼n=N(k)) Doppler processing, waveform and performance measures
for some pulsed Doppler and MTD radars.
+ (1 ¡ n2s )(¼Tr (k))2 =2: Ortung und Navigation, 3 (1981), 417—457.
750 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 33, NO. 3 JULY 1997
André Ferrari (S’91–M’92) was born in Senlis, France, on June 2, 1964. He
received the engineer degree from École Centrale de Lyon in 1988 and the DEA
degree and Ph.D. degree from University of Nice Sophia Antipolis in 1989 and
1992, respectively.
He is currently an Associate Professor at University of Nice Sophia Antipolis
and member of I3S laboratory. His research interest include spectral estimation
and statistical signal processing with applications to radar.