SOME RADAR TOPICS:

WAVEFORM DESIGN, RANGE CFAR

AND TARGET RECOGNITION
H. Rohling
Technical University Hamburg-Harburg
Hamburg, Germany
[email protected]
Abstract The rst RADAR patent was applied for by Christian Huelsmeyer on
April 30, 1904 at the patent oce in Berlin, Germany. He was moti-
vated by a ship accident on the river Weser and called his experimental
system Telemobiloscope. In this chapter some important and modern
topics in radar system design and radar signal processing will be dis-
cussed. Waveform design is one innovative topic where new results are
available for special applications like automotive radar. Detection the-
ory is a fundamental radar topic which will be discussed in this chapter
for new range CFAR schemes which are essential for all radar systems.
Target recognition has for many years been the dream of all radar en-
gineers. New results for target classication will be discussed for some
automotive radar sensors.
Keywords:
automotive radar; continuous wave radar; CFAR processing; rank-
order; lters; FSK modulation; matched lters; unmatched lters.
1. Introduction
The objective of this chapter is to discuss some important contri-
butions to radar system design and digital radar signal processing. The
focus is on waveform design in general and on automotive applications in
particular. Target detection is an important issue for all radar systems.
Therefore some range CFAR (constant false alarm rate) procedures will
be discussed which can be applied, especially in multiple target situa-
tions, to avoid masking. Additionally some new results are discussed for
target recognition systems which have been developed for automotive
applications.
2
2. Combination of LFMCW and FSK
modulation principles for automotive radar
systems
High performance automotive radar systems are currently under de-
velopment for various applications. Comfort systems like Adaptive Cruise
Control (ACC) are already available on the market as 77 GHz radars.
Target range and velocity are measured simultaneously with high resolu-
tion and accuracy even in multi-target situations, but the measurement
and processing time to detect the relevant object is approximately 100
ms. Future developments will be more concentrated on safety applica-
tions like Collision Avoidance (CA) or Autonomous Driving (AD). In
this case the requirements for reliability (extreme low false alarm rate)
and reaction time (extreme short delay) are much higher compared with
ACC systems.
To meet all these system requirements specic waveform design tech-
niques must be considered. For ACC systems both radar types of classi-
cal pulse waveform with ultra short pulse length (10 ns) or alternatively
continuous wave (CW) transmit signal with a bandwidth of 150 MHz
are considered. The main advantage of CW systems in comparison with
classical pulse waveforms is the low measurement time and low compu-
tational complexity.
This section describes a new waveform design for automotive appli-
cations based on CW transmit signals which lead to an extremely short
measurement time. The basic idea is a combination of linear frequency
modulation (LFM) and FSK CW waveforms in an intertwined technique.
Unambiguous range and velocity measurement with high resolution and
accuracy can be required in this case even in multi-target situations.
After introductions to FSK and LFM waveform design techniques in
sections 2 and 1 the combined and intertwined waveform is described in
detail in section 1.
PURE FSK MODULATION
Pure FSK modulation (as shown in Figure 1 (a)) uses two discrete
frequencies f
A
and f
B
(so-called two frequency measurement) [1] in the
transmit signal. Each frequency is transmitted inside a so-called co-
herent processing interval (CPI) of length T
CPI
(e.g. T
CPI
= 5 ms).
Using a homodyne receiver the echo signal is down converted by the
instantaneous frequency into base band and sampled N times. The fre-
quency step f
Step
= f
B
f
A
is small and will be chosen to depend on the
maximum unambiguous target range. The time-discrete receive signal
is Fourier transformed in each CPI of length T
CPI
and targets will be
Waveform Design, Range CFAR and Target Recognition 3
detected by an amplitude threshold (CFAR). Due to the small frequency
step a single target will be detected at the same Doppler frequency po-
sition in the adjacent CPIs but with dierent phase information on the
two spectral peaks. The phase dierence =
B

A
in the complex
spectra is the basis for the target range (R) estimation. The relation be-
tween the target distance and phase dierence is given by the equation:
R =
c
4 f
Step
. (1)
To achieve an unambiguous maximum range measurement of 150 m
a frequency step of f
Step
=1 MHz is necessary. In this case the target
resolution only depends on the CPI length T
CPI
. The technically sim-
ple VCO modulation is an additional advantage of this waveform. But
this FSK waveform does not allow any target resolution in the range
direction, which is an important disadvantage of this measurement tech-
nique. Especially in the automotive trac environment, more than a
single xed target will occur simultaneously inside an antenna beam.
These xed targets cannot be resolved by a FSK waveform.
CPI 0 CPI 1 Chirp 0 Chirp 1
(a) (b)
) (J B
6
5JAF
B
) (J B
6
5MAAF
B J J
+21
6
+21
6 2
0
+DEHF
6
+DEHF
6 2
0
Figure 1. Two CW waveform principles: (a) FSK modulation, (b) LFM
PURE LINEAR FREQUENCY MODULATION
Radars which use a pure LFM technique modulate the transmit fre-
quency with a triangular waveform [6]. The oscillator sweep is given
by f
Sweep
. A typical value for the bandwidth is f
Sweep
=150 MHz to
achieve a range resolution of R =
c
2f
Sweep
= 1 m. In general, a single
sweep of the LFM waveform gives an ambiguous measurement in range
and relative velocity. The down converted receive signal is sampled and
Fourier transformed inside a single CPI. If a spectral peak is detected in
the Fourier spectrum at index (normalized integer frequency) the am-
biguities in target range and velocity can be described in an R-v-diagram
by the equation:
4
=
v
v

R
R

v
v
=
R
R
+ (2)
where v gives the velocity resolution resulting from the CPI duration
T
Chirp
( v =

2T
Chirp
= 0.8 m/s, is the wavelength of 4 mm @ 77 GHz
and T
Chirp
= 2.5 ms ).
Due to resulting range-velocity ambiguities, further measurements
with dierent chirp gradients in the waveform are necessary to achieve
an unambiguous range-velocity measurement, even in multi-target sit-
uations. The well-known up/down-chirp principle as it is depicted in
Figure 1 (b) is described in detail in [5]. LFM waveforms can be used
even in multi-target environments, but the extended measurement time
is an important drawback of this LFM technique.
COMBINED FSK AND LFM WAVEFORMS
The combination of FSK and LFM waveform design principles oers
the possibility of unambiguous target range and velocity measurement
simultaneously. The transmit waveform consists in this case of two linear
frequency modulated up-chirp signals (the intertwined signal sequences
are called A and B). The two chirp signals will be transmitted in an
intertwined sequence (ABABAB...), where the stepwise frequency mod-
ulated sequence A is used as a reference signal while the second up-chirp
signal is shifted in frequency by f
Shift
. The received signal is down con-
verted into base band and directly sampled at the end of each frequency
step. The combined and intertwined waveform concept is depicted in
Figure 2.
A
B
A
B
A
B
t
0
) (t f
T
B T
f
,
A T
f
,
Shift
f
1
=
N
f
f
Sweep
Incr
Sweep
f
Chirp
T
Figure 2. Combined FSK-LFMCW waveform principle.
Each signal sequence A or B will be processed separately by using
the Fourier transform and CFAR target detection techniques. A single
target with specic range and velocity will be detected in both sequences
Waveform Design, Range CFAR and Target Recognition 5
at the same integer index =
A
=
B
in the FFT-output signal of the
two processed spectra. In each signal sequence A or B the same target
range and velocity ambiguities will occur as described in equation 2. But
the measured phases
A
and
B
of the two (complex) spectral peaks are
dierent and include the ne target range and velocity information which
can be used for ambiguity resolution. Due to the coherent measurement
technique in sequences A, B the phase dierence =
B

A
can be
evaluated for target range and velocity estimation. The measured phase
dierence can be described analytically by the following equation:
f
r
e
q
u
e
n
c
y
m
e
a
s
u
r
e
m
e
n
t
Target
Detection!
phase
measurement
0
v
0
v
0
R R
>
Shift
f
>
Sweep
f
Figure 3. Graphical resolution principle of ambiguous frequency and phase mea-
surements.
=

N 1

v
v
4 R
f
Shift
c
(3)
where N is the number of frequency steps (or receive signal samples) in
each transmit signal sequence A, B. In this rst step is ambiguous
but it is possible to resolve these ambiguities by combining the two
measurement results of equations 2 and 3. The intersection point of the
two measurement results is shown in Figure 3 in a graphical way. The
analysis leads to an unambiguous target range R
0
and relative velocity
v
0
:
R
0
=
c R

(N 1)
c 4 (N 1) f
Shift
R
(4)
v
0
=
(N 1) v

c 4 f
Shift
R
c 4 (N 1) f
Shift
R
(5)
This new intertwined waveform shows that unambiguous target range
and velocity measurements are possible even in a multi-target environ-
ment. An important advantage is the short measurement and processing
time.
6
SYSTEM EXAMPLE
In this section a waveform design based on the new intertwined signal
is developed as an example for automotive applications. The signal
bandwidth is f
Sweep
=150 MHz to fulll the range resolution requirement
of 1 m. The stepwise frequency modulation is split into N = 256 separate
bursts of f
Incr
=
150MHz
255
= 588 kHz each. The measurement time inside
a single burst A or B is assumed to be 5 s resulting in a chirp duration
of the intertwined signal of T
Chirp
= 2.56 ms. This results in a velocity
resolution of v =

2T
Chirp
= 2.7 km/h.
The important waveform parameter f
Shift
is optimized on the basis
of high range and velocity accuracy. The highest accuracy occurs if the
intersection point in the R-v-diagram results from two orthogonal lines.
For this reason the frequency shift between the signal sequences A
and B is f
Shift
=
1
2
f
Incr
= 294 kHz.
In this specic case equations 4 and 5 turn into
R
0
R
=
N 1
2


2
(6)
v
0
v
=
N 1
2
+

2
(7)
The proposed intertwined and stepwise CW waveforms show high
performance in simultaneous range and velocity measurement accuracy.
The main advantage is the short measurement time in comparison to
classical LFM waveforms while the resolution and accuracy are un-
changed. The properties of the new intertwined CW waveform technique
are quite promising. This concept is a good basis for high performance
automotive radar systems with dierent safety applications (e.g. pre
crash) which require ultra short measurement and processing times.
3. Automotive Radar Network Based On 77GHz
FMCW Sensors
Automotive radar systems need to have the capability to measure
range, velocity and azimuth angle simultaneously for all point and ex-
tended targets inside the observation area. Short measurement time
even in dense target situations, and high range accuracy and resolution,
are required in all automotive applications. We will distinguish in this
section between single radar sensors and radar network systems. So-
called far distance single radar sensors use an observation area of 10

in azimuth angle and a maximum range of up to 200m.

Waveform Design, Range CFAR and Target Recognition 7
If, in contrast, a large azimuth angle coverage and a short maximum
range are required, radar network systems with e.g. four individual sen-
sors mounted behind the front bumper are used instead of a single radar
sensor. Typical automotive applications for radar networks with a large
azimuth angle coverage but limited range are, for example, Collision
Avoidance and Pre Crash Warning. Figure 4 illustrates the observation
area considered in this section. The relevant system parameters for a
short-range radar network are given in Table 1.
The developed near distance single radar sensor provides target range
and radial velocity simultaneously with high accuracy and for all objects
inside the observation area. It is characteristic of radar networks signal
processing that the angular position of each target is calculated by means
of multilateration techniques based on the sensor specic measured tar-
get ranges inside the network [7, 8]. This is to derive the desired target
position by calculating the intersection point of all range measurements
from dierent radar sensor positions.
Figure 4. Observation area of the radar network system.
Table 1. Requirements for a single sensor in a radar network system
Parameter Value
Observation area 120

in azimuth
Maximum range 40m
Range resolution 0.4m
Range accuracy (required by multilateration) 0.02m
Velocity resolution 1m/s
Velocity accuracy 0.3m/s
Target acquisition time 20-100ms
8
RADAR NETWORK
Architecture. In this section we present a short-range radar net-
work that consists of four distributed 77GHz radar sensors [9]. All sen-
sors are mounted behind the front bumper of the vehicle in an invisible
form. In contrast to already known radar networks in the 24GHz fre-
quency domain, the presented sensors are working at a 77GHz carrier
frequency. Due to European regulations, this frequency band provides
a wider bandwidth for automotive radar applications. To achieve the
desired high range accuracy of 2cm, which is required by the multilater-
ation process, a frequency sweep of 1GHz has to be utilized within the
Linear Frequency Modulated Continuous Waveform (LFMCW). Figure
5 shows an image of the 77GHz prototype sensor [10]. Time synchroniza-
tion between the individual sensors is needed to avoid any interference
situations between the radar sensors. With additional carrier synchro-
nization between the sensors the radar network could even provide the
capability of bistatic measurements. The measurement results described
in this section are based on monostatic sensor measurements.
Figure 5. Close-up of a single 77GHz radar sensor.
Experimental system. To validate the analytical results some ex-
perimental cars, shown in Figure 6, have been equipped with a radar
network. These test cars are used for data acquisition and recording to
optimize the signal processing algorithms. In section 9 experimental re-
sults are presented to illustrate the theoretical results of developed and
applied algorithms.
SINGLE SENSOR SIGNAL PROCESSING
For data acquisition a classical linear FMCW waveform is used which
consists of four individual chirp signals (see Figure 7) and covers a band-
width of 1GHz, which corresponds to a range resolution of 0.4m under
realistic signal processing conditions. This waveform combines high ac-
Waveform Design, Range CFAR and Target Recognition 9
Figure 6. Two experimental cars equipped with 77GHz radar sensors.
curacy measurements in range and velocity and reliable target detection
in multi-target situations. Furthermore, FMCW signal processing tech-
niques compared with pulse radar waveforms lead to reduced computa-
tional complexity and hardware requirements [11].
Four individual chirps provide sucient redundancy in multi-target
or extended target situations [4] to suppress ghost targets in the range-
velocity processing. For each individual chirp signal the beat frequencies
df
1,
... df
4
, shown in Figure 7, will be estimated by applying an FFT.
Figure 7. FMCW radar waveform (red) and the corresponding receive signal (blue)
for a single target situation.
Table 2 gives a summary of the considered waveform parameters. It
is a real technical challenge to handle, in multiple and extended target
situations, the large number of detections in the data association, track-
ing as well as range, velocity and angle parameter estimation procedure.
Therefore, a specic signal processing technique has been developed and
is implemented in the experimental system.
The classical FMCW radar signal processing scheme is structured into
the following dierent independent blocks: beat frequency estimation
based on FFT, target detection (CFAR), range and velocity processing,
multilateration for azimuth angle measurement and tracking; see Figure
8 and [12]. Even as an enhancement of classical signal processing a
classication procedure could be used to derive additional information
about the target object class [13].
10
Table 2. Waveform Parameters
Parameter Value
Center frequency 77GHz
Number of Chirps 4
Single chip duration 2 ms
First sweep bandwidth 1GHz
Second sweep bandwidth 500MHz
Multilateration Tracking
A
s
s
o
c
i
a
t
i
o
n
A
s
s
o
c
i
a
t
i
o
n
Range-
Velocity
Estimation
A
s
s
o
c
i
a
t
i
o
n
Detection FFT
Figure 8. Classical FMCW radar signal processing
After CFAR detection each signal-processing block contains an inde-
pendent association procedure to combine measurements from dierent
chirps and dierent radar sensors, which belong to a single point or
even extended target. In the following, all signal processing steps are
presented and will be discussed in detail.
The objective of this section is to optimize the classical signal process-
ing structure especially for automotive applications. The main draw-
back of the classical signal processing structure is that in multiple and
extended target situations each of the three independent association pro-
cedures (Figure 8) will induce some ghost targets. This network behav-
ior has been observed in many sets of measured data. Based on these
observations and results inside a classical radar network an alternative
signal processing structure is proposed herein, which is based on a joint
optimization of the three independent association schemes.
Target detection. Due to the large angular coverage considered it
is characteristic for automotive applications that each radar sensor has
many detections at the CFAR procedure output due to multiple and
extended target situations. For reliable target detection in this multi
target environment, an ordered-statistic (OS) constant-false-alarm-rate
(CFAR) thresholding has been applied which showed the best experi-
mental results [14].
As a result of the FMCW waveform with four individual chirp signals,
a proper detection process results in four detected beat frequencies f
C,S
per point target and sensor. In total a single target, detected by the
Waveform Design, Range CFAR and Target Recognition 11
radar network, will lead to 16 detected beat frequencies at the FFT out-
put. The vector

m
f
combines these beat frequencies and therefore de-
scribes all information which is available in the signal-processing scheme
for each individual target.
m
f
=
_
_
f
1,1
, f
2,1
, f
3,1
, f
4,1
. .
Sensor1
, , f
1,4
, f
2,4
, f
3,4
, f
4,4
. .
Sensor4
_
_
T
(8)
Range-Velocity Estimation. Each measured beat frequency con-
tains information about the target range and velocity in an ambiguous
way. But each individual target with range r
S
and velocity v
S
leads to a
deterministic beat frequency for each chirp signal of the waveform. This
beat frequency and the relation to target range and velocity is given by
the linear equation:
f
C,S
= a
C
r
S
+b
C
v
S
. (9)
Parameters a
C
and b
C
depend on chirp characteristics like chirp du-
ration, bandwidth and carrier frequency [6]. Based on the 4 beat fre-
quencies measured by a single sensor the point target range and velocity
can be derived simultaneously by an intersection process. In this case
and in a single point target situation the four measured frequencies are
transformed into target range and velocity, unambiguously. But in mul-
tiple or even extended target situations this range velocity calculation
could lead to some ghost targets.
Each sensor of the radar network has an individual position behind the
front bumper. Therefore, each sensor will calculate individual values for
target range and velocity based on the four measured beat frequencies,
equation 8, inside the FMCW waveform. The measurement result is
described by an eight-element parameter vector.
m
t
=
_
_
r
1
, v
1
. .
Sensor1
, , r
4
, v
4
. .
Sensor4
_
_
T
(10)
A set of linear equations can now be derived which describes the
relation between 16 measured beat frequencies and sensor specic target
range and velocity parameters.
m
f
= C m
t
(11)
In multi-target situations the association of the detected beat frequen-
cies at the FFT output to dierent targets is not trivial. Therefore, a
12
data association process has to be performed using the redundancy given
by the four chirp signals inside a single waveform.
RADAR NETWORK SIGNAL PROCESSING
The objective of radar network signal processing is to calculate the
azimuth angle (or target position in Cartesian coordinates) of each target
in multiple or even extended target situations based on the precise range
measurement of each radar sensor. Furthermore the tracking procedure
is part of the network processing.
Multilateration Procedure. To derive target position and velocity
described by a target state vector in Cartesian coordinates

t = (t
x
, t
y
, v
x
, v
y
)
T
(12)
a multilateration technique is used. Based on sensor specic range and
velocity measurements for each individual point target the state vector
can be estimated if the position of each sensor behind the front bumper
is taken into account, which is described by the vector:
s = (s
x
, s
y
)
T
. (13)
For each sensor the target range, see (9) , can be calculated by a
nonlinear equation if target and sensor Cartesian positions are known.
r
S
=
_
(t
x
s
x
)
2
+ (t
y
s
y
)
2
. (14)
The target radial velocity in (9) can be processed as follows:
v
S
=
t
x
s
x
r
S
v
x
+
t
y
s
y
r
S
v
y
. (15)
Combining both equations, the relation between target state vector

t
and the measurement parameter vector m
t
, equation 10, can be formu-
lated by the nonlinear equation
m
t
= h(

t). (16)
The Jacobian matrix
H

t
0
=
h(

t)

t
0
(17)
is used in an iterative Gauss-Newton algorithm to estimate the target
position and velocity in Cartesian coordinates based on the given mea-
surements and a given initial position estimate [15].
Waveform Design, Range CFAR and Target Recognition 13
At the beginning of the multilateration procedure an association pro-
cess is integrated to select the sensor specic range and velocity measure-
ments belonging to a single target. In multiple and extended target sit-
uations there is high risk of ghost targets due to errors in the association
procedure. Furthermore, in situations with low target detection proba-
bility the multilateration procedure does not have sucient information
for calculating the target state vector. While the data association for
range-velocity processing was provided by four individual chirp signals
within the waveform, the data association for multilateration processing
is provided by measurements of four individual radar sensors at dierent
positions.
Target Tracking. Normally radar tracking is based purely on plot-
to-track association. In this section we develop a radar network and
FMCW waveform specic frequency-to-track association as an extension
of the classical tracking procedure. For pulse radar networks a similar
idea of range-to-track association schemes have been published in [12].
In this case the multilateration technique is not processed independently
and explicitly but is integrated into the tracking procedure which is
based on Kalman ltering.
For automotive applications with relative low velocities and a high
update rate 1/T, a pure linear motion model with constant velocity can
be considered. The respective state transition matrix A for a constant-
velocity trajectory can be used to calculate the predicted target state
vector for the next time step by the following equation:

t
k+1
= A

t
k
+ w
k
=
_

_
1 T
1 T
1
1
_

_
t
x
t
y
v
x
v
y
_

_
+
_

_
0
0
w
x
w
y
_

_
. (18)
Here

t
k
is the target state vector at time index k and w
k
contains two
random variables which describe the unknown process error, which is
assumed to be a Gaussian random variable with expectation zero and
covariance matrix Q. In addition to the target dynamic model, a mea-
surement equation is needed to implement the Kalman lter. This mea-
surement equation maps the state vector

t
k
to the measurement domain.
In the next section dierent measurement equations are considered to
handle various types of association strategies.
14
JOINTLY OPTIMIZED RADAR NETWORK
SIGNAL PROCESSING
In general, the tracking procedure starts with an association process
to combine the established track parameter with the radar sensor or
radar network measurements. Errors in the association process will al-
ways lead to ghost targets. But the general requirement for automotive
applications is to keep the false alarm probability as low as possible,
which underlines the importance of the association process for radar
networks.
There are several possibilities in radar networks to design this associ-
ation process in the tracking procedure:
The target state vector

t
k
measured by the multilateration pro-
cedure can be considered directly as a target plot input of the
association process. In this case, the input of the Kalman lter
describes the same parameters that the internal state vector does.
It is characteristic for the plot-to-track association procedure that
the measurement equation contains directly the target state vector

t
k
which is inuenced by noise n
s
k
only:
y
s
k
= [t
x
, t
y
, v
x
, v
y
]
T
k
+n
s
k
=

t
k
+n
s
k
. (19)
Alternatively the radar sensor specic measured ranges and veloc-
ities m
t
k
can be used for a track update. In this case the tracking
procedure can even be applied in the low target detection situa-
tion where the multilateration process cannot be applied. In the
range-velocity-to-track association scheme the corresponding mea-
surement equation is based on range and velocity calculations and
has a nonlinear analytical structure,
y
t
k
= [r
1
, v
1
, , r
4
, v
4
]
T
k
+n
t
k
= m
t
k
+n
t
k
(20)
= h(

t
k
) +n
t
k
.
It has to be noted that the measurement values for range and ve-
locity are not uncorrelated according to the LFMCW measurement
described in section 8. As a consequence, the observed measure-
ment errors n
t
k
can also be considered as correlated random vari-
ables for a single sensors data. For 24GHz pulse radar networks,
developed also for automotive applications, a similar idea has been
described by a range-to-track association scheme [12], because no
velocity measurements are provided in such a radar network.
Waveform Design, Range CFAR and Target Recognition 15
Finally for an FMCW radar waveform all the measured single beat
frequencies m
f
can be used directly for the association process and
track update. This technique will be called frequency-to-track as-
sociation. In this case each radar detection and beat frequency
measurement can be directly integrated into the tracking process.
An explicit and independent calculation of the range-velocity pa-
rameter and multilateration processing is not necessary in this
case. This joint association process reduces the ghost target prob-
ability dramatically and improves the radar network performance.
y
f
k
= [f
1,1
, f
2,1
, , f
4,1
, , f
4,4
]
T
k
+n
f
k
= m
f
k
+n
f
k
= C m
t
k
+n
f
k
(21)
= C h(

t
k
) +n
f
k
The vector n
f
k
describes the unknown additive measurement noise,
which is assumed in accordance with Kalman lter theory to be a
Gaussian random variable with zero mean and covariance matrix
R. Instead of the additive noise termn
t
k
in equation (20), the errors
of the dierent measurement values are assumed to be statistically
independent and identically Gaussian distributed, so
R = E{n
f
k
n
fT
k
} =
2
f
I. (22)
The term E{} denotes the expected value and I is the identity
matrix. This covariance matrix can be derived from the radar
sensor characteristics.
The respective Kalman lter equations for the position correction and
prediction steps can now be formulated based on equations (18) and (19),
(20) or (21) accordingly for the dierent mentioned association schemes.
Since the measurement equation is nonlinear in case of range-velocity-
to-track or frequency-to-track association, the Extended Kalman lter
is used for this particular application [16].
In Figure 9 a block diagram of the frequency-to-track processing is
given. The association procedures are no longer processed stepwise
at three dierent places in the block diagram compared to the general
classical radar network scheme described in Figure 8.
Even a small subset of the maximum possible sixteen beat frequencies
is sucient for track update processing based on the frequency-to-track
association scheme. Almost all association errors could be avoided in
multiple and extended target situations applying this procedure. This
16
Tracking Multilateration
Range-
Velocity
Estimation
A
s
s
o
c
i
a
t
i
o
n
Detection FFT
Figure 9. Block diagram of frequency-to-track association
proposed joint optimization procedure shows increased performance and
is additionally quite robust in all situations when some radar sensors
in the radar network have low detection probability. The ghost target
probability is dramatically reduced.
EXPERIMENTAL RESULTS
As already mentioned, the TUHH experimental car has been equipped
with the described 77GHz radar network to validate and prove the e-
ciency of the derived algorithms. This radar network was used to record
measurement data of typical scenarios in real street applications. Based
on these recorded data, a comparison of the dierent signal processing
strategies (classical or jointly) for radar network signal processing has
been performed.
Typical targets for automotive radar networks are moving cars inside
the observation area. Compared with the single radar sensor range reso-
lution of 0.4m a common car cannot be considered any longer as a point
but as an extended target. Therefore, each sensor will measure several
echo signals in dierent but closely related range gates for this single
car.
Measurements of such extended targets contain and describe all signal
processing and association eects discussed in the previous sections.
In the classical signal processing case the target azimuth angle is cal-
culated in the radar network based on multilateration techniques. In
this case an extremely high range accuracy of 2cm is required due to
the small baseline of radar sensor position inside the network and the
required position accuracy. Furthermore, all sensors are observing the
car from slightly dierent aspect angles and can therefore detect dier-
ent reection centers. It is obvious that the data association technique
becomes very crucial in such situations and the risk of producing ghost
targets caused by multiple detections and misassigned measurements in
the three independent association schemes is rather high.
On the other hand, the high update rate compared to realistic veloci-
ties and accelerations of cars results in a high quality target prediction in
the tracking procedure. Therefore, nearest neighbour and gating tech-
Waveform Design, Range CFAR and Target Recognition 17
Figure 10. Range-velocity-to-track association with a range gate size of 0.5m and
a velocity gate size of 1m/s (left) and frequency-to-track association with a gate size
of 1FFT bin (right)
niques with rather small gate sizes can be used for signal processing and
association procedures.
As an example, the measurement results of a car driving straight to-
wards the radar-sensor network with a velocity of approximately 5m/s
have been analyzed. Figure 10 shows the results of the dierent sig-
nal processing stages. The car can be considered from the single radar
sensor point of view as an extended target, which leads to many echo
signals with dierent ranges. Therefore the three association steps in the
classical signal processing procedure must be considered quite carefully.
It can be seen that the target positions derived by a pure multilater-
ation procedure (green circles) in the classical signal processing scheme
have poor quality and accuracy due to many association errors in the
range-velocity measurement inside each sensor and inside the multilat-
eration step. Considering these green dots it is hopeless to establish a
plot-to-track association in the tracking procedure.
The situation can be improved by a range-to-track association tech-
nique. The blue line in the left part of Figure 10 shows the result of a
range-velocity-to track association procedure as described in section 8.
In this case the target can be tracked over the complete measurement
time but with limited accuracy, especially in the azimuth angle. The
target position accuracy is increased for short-range positions.
The right part of Figure 10 shows the result of the joint optimiza-
tion procedure based on a direct frequency-to-track association as de-
18
scribed in section 8. The green circles again show the results of a pure
multilateration process as in the left part for the extended target mea-
surement situation. The most obvious dierence is improved accuracy
in angular estimation of the target position. From Figure 10, the im-
proved performance of the proposed joint optimization process and the
frequency-to-track association can be seen.
4. Range CFAR Techniques
The general task of primary radars used in air or vessel trac control
is to detect all targets inside the observation area and to estimate their
range, azimuth and radial velocity parameters respectively. The target
detection scheme would be an easy task if the echo signal was observed
before an empty or statistically completely known noise or clutter signal
background. In this case all received echo signal amplitudes would be
compared with a xed threshold, which is based on the noise and clutter
statistic only, and targets are detected in all cases when this threshold
is exceeded by the echo signal inside the test cell.
But in real radar applications many dierent noise and clutter back-
ground signal situations can occur. The target echo signal practically
always appears before a background signal, which is lled with point,
area or even extended clutter and additional superimposed noise. Fur-
thermore the location of this background clutter varies in time, position
and intensity. Clutter is, in real applications, a complicated time and
space variant stochastic process.
All these conditions call for an adaptive procedure in detection and
signal processing, operating not with a xed but with a variable thresh-
old in the detection procedure, to be determined in accordance with
the locally observed clutter situation with dierent range extension, in-
tensity and uctuation. In a rst step of the detection procedure the
unknown parameter of a certain statistical background signal is always
estimated by analysing the signal inside a xed window size, which is ori-
ented in the range direction surrounding the radar test cell. The general
detection procedure is shown in a block diagram in Figure 11 where the
sliding range window is split into two parts, the leading and lagging part
surrounding the test cell. Additionally some guard cells are introduced
to reduce self-interferences in a real target echo situation.
All data inside the window will be used to estimate the unknown
statistical parameters of the background clutter and to calculate the
adaptive threshold for target detection. All the background signals,
undesired as they are from the standpoint of detection and tracking, are
denoted just as clutter. The detection procedure has to distinguish
Waveform Design, Range CFAR and Target Recognition 19
Figure 11. General architecture of range CFAR procedures
between useful target echoes and all possible clutter situations. Clutter
is not just a uniformly distributed sequence of random variables but can
be caused in practical applications by a number of dierent physical
sources.
Therefore, the length of the sliding window will be chosen as a com-
promise based on rough knowledge about the typical clutter extension.
To get good estimation performance (low variance) in a homogeneous
clutter environment the window size N should be as large as possible.
But the window length N must be adapted to the typical range exten-
sion of homogeneous clutter areas to fulll the statistical requirement
of identically distributed clutter random variables. In typical air trac
control radars the number of range cells is e.g. between N=16 and 32.
Radar target detection in noise
In a rst simple model for target detection it is assumed that the
background clutter can be described by a statistical model in which
the dierent range cells inside the sliding window contain statistically
independent identically exponentially distributed (iid) random variables
{X
1
,. . . ,X
N
}. The probability density function (pdf) of exponentially
distributed clutter variables is fully described by the equation:
p
0
(x) = 1/ e
x/
, x 0. (23)
20
In this ideal case it is assumed that is a known parameter which
describes the expected value of the exponentially distributed random
variables. The variance of the random variable X in this case is
2
=

2
. The false alarm probability (P
fa
) depends on the noise and clutter
statistic only and a certain threshold S will be calculated analytically
as follows:
P
fa
=

_
S
p(x)dx (24)
S = T . (25)
The factor T of the threshold S can be described for a given false
alarm probability (P
fa
) in this case analytically by the equation:
T = ln
1
P
fa
. (26)
The non-uctuating target amplitude statistic can be described by
the Rician pdf:
p
1
(x) =
_
_
_
x

2
0
e

x
2
+c
2
2
2
0
I
0
_
xc

2
0
_
, x 0;
0, otherwise.
(27)
Therefore the detection probability (P
d
) for a non-uctuating target
in homogeneous noise is:
P
d
=

_
T
p
1
(x)dx. (28)
The general objective of all radar detection procedures is to get a
constant false alarm rate (CFAR) due to the fact that the test cell almost
always contains clutter and noise and only in a very few cases contains
radar target echo signals. The statistical model and general detection
procedure, in which the detector is xed only with regard to the noise
and clutter statistic and independently to the target statistic, has been
developed by Neyman and Pearson.
But in real radar applications the average noise and clutter power
level () is unknown and must be estimated in the detection procedure
rst. This is done by several published CFAR procedures, which will be
discussed in this section, where each specic CFAR technique is moti-
vated by assumptions about a specic background signal or target signal
model.
Waveform Design, Range CFAR and Target Recognition 21
Range CFAR procedures with sliding window techniques
Each developed and published CFAR technique refers implicitly to a
certain background clutter or even target model. Therefore, in the fol-
lowing these assumptions will be described explicitly and the dierent
CFAR procedures considered will be compared in some clutter situa-
tions. The amplitude in each individual range cell is tested. Therefore
a window of xed size is applied to each range cell inside the full range
coverage using a sliding technique. The amplitudes in the leading and
lagging reference cells are used in a signal processing procedure to esti-
mate the unknown statistical parameters of the clutter background sig-
nal. Based on these estimated parameters the detection threshold will
be calculated. The dierent range CFAR techniques dier in the way to
estimate the statistical parameters. In the following the background sig-
nal model and the resulting motivation for each CFAR technique will be
described. Furthermore each range CFAR technique will be applied in
four dierent but (for radar applications) characteristic signal situations
which consist of: pure noise; local clutter; single target in noise; and two
targets respectively. From these examples the characteristic behavior of
each CFAR procedure can be seen clearly.
Cell averaging CA CFAR. In this rst signal model it is assumed
that the clutter and noise background at the output of a square law
detector can be described by statistically independent and identically
distributed (iid) exponential random variables with a single exception:
the average clutter plus noise power level is unknown. The optimised
signal processing technique in this situation, from a statistical point
of view, is to calculate an estimation of the clutter power level just
by applying the arithmetic mean to the received amplitudes inside the
considered window.
Z =
_
1
N
N

i=1
X
i
_
. (29)
The arithmetic mean Z has excellent estimation performance. The esti-
mation Z is unbiased, which means
E
_
1
N
N

i=1
X
i
_
=
1
N
N

i=1
E [X
i
] =
and shows additionally a minimum estimation variance.
V ar
_
1
N
N

i=1
X
i
_
=
1
N
2
N

i=1
V ar [X
i
] =
1
N

2
. (30)
22
The false alarm rate is given by [4] as:
P
fa
= P(Y T
CA
Z) = (1 +T
CA
)
N
. (31)
If this estimation procedure is applied to the random variables in-
side the range window this CFAR procedure is called cell averaging,
CA-CFAR. The statistical performance is excellent if the assumptions
of homogeneous clutter inside the reference window are fullled in the
statistical model and in the real world application. It is not clear to
the author who rst analyzed and published this CA-CFAR idea, but
Nitzberg [17] published a paper in 1978 analyzing CA-CFAR for uc-
tuating targets. To demonstrate the general CFAR characteristic some
typical signal situations are generated which are considered to be char-
acteristic for radar applications. These are: a pure noise background
signal inside the full range coverage; a local clutter area over 10 adja-
cent range gates superimposed with noise; a single target and a double
target situation (20 dB SNR each) superimposed with noise. The clut-
ter power was chosen 13 dB above noise power. Figure 12 shows the
resulting threshold S in these noise, clutter and target situations when
the CA-CFAR procedure is applied.
Figure 12. CA-CFAR (N=24), P
fa
= 10
6
Figure 12 shows the appropriately calculated detection threshold in
the pure noise situation (12 a). This homogeneous noise model with
unknown noise power was the motivation to develop CA-CFAR. The
detection procedure adapts quite well in the local clutter with some
losses (increased P
fa
) at the clutter edges (12 b). The characteristic
behavior in target situations is not acceptable (12 c). In the two target
Waveform Design, Range CFAR and Target Recognition 23
situation both targets are not even detected by CA-CFAR due to the
resulting masking situation (12 d).
Cell averaging with greatest of CAGO-CFAR. The clutter and
noise signals are varying in time and position and the average clutter
power level can uctuate in dierent range areas and range cells. If a
CA-CFAR is applied in the radar detector it may happen that the sliding
window is located in the transition between a pure noise and strong
clutter area with dierent average power level, as shown in Figure 12c
for example. From a statistical point of view this means that the random
variables inside the sliding window are no longer identically distributed
but have dierent expected values in their individual statistics. CA-
CFAR leads in such cases to an increase in false alarm rate (P
fa
), which
is unacceptable for practical applications and requirements.
Therefore the clutter model is extended and non-homogeneous clut-
ter situations are integrated, such as in typical transition areas between
noise and beginning clutter areas. Referring to such realistic clutter sit-
uations a new CFAR procedure has been designed by Vilhelm G. Hansen
in his well-known paper [18]. In order to demonstrate the advantage of
this CFAR technique it is important to recall the CA-CFAR procedure.
In this case the arithmetic mean is calculated in the leading and lagging
part of the range window and both values are summed up; see Figure
11.
Figure 13. CAGO-CFAR (N=24), P
fa
= 10
6
In the extended CFAR case a cell averaging technique is used in each
part of the range window but with a greatest of selection (CAGO-
CFAR).
24
Z = Max
_
_
_
_
2
N
N/2

i=1
X
i
_
_
,
_
_
2
N
N

i=N/2+1
X
i
_
_
_
_
. (32)
If CAGO-CFAR is applied in such typical clutter situations the false
alarm rate is reduced at the clutter edges but the detection rate is si-
multaneously reduced slightly, see Figure 13.
CAGO-CFAR shows a clear advantage in typical transition areas be-
tween pure noise and strong clutter or between two dierent clutter re-
gions. But it simultaneously reduces the sensitivity in target detection
in a homogeneous clutter situation. This is a compromise, which always
occurs in radar detection due to the variety of dierent noise and clutter
background situations, which can occur in real radar applications. It is
a useful extension of CA-CFAR but shows the same characteristics in
multiple target situations.
Moshe Weiss [25] developed an extension of CAGO-CFAR to get bet-
ter performance in multiple target situations and designed cell averaging
CFAR with smallest of (CASO-CFAR) procedures and analysed the per-
formance especially in multiple target situations.
These CFAR procedures suer from the fact that they are specically
tailored to the assumption of uniform and homogeneous clutter inside
the reference window. Based on these assumptions, they estimate the
unknown clutter power level using the unbiased and most ecient arith-
metic mean estimator. Improved CFAR procedures should be robust
with respect to dierent clutter background and target situations. Also
in non-homogeneous situations CFAR techniques should remain able to
provide reliable clutter power estimations.
OS-CFAR. Applying CAGO-CFAR in the detection procedure brings
several advantages in clutter transition areas. But the CA and CAGO-
CFAR detection procedures behave very sensitively in multiple target
situations and show pure performance. This observation has been de-
scribed in [14]. It was shown that even in a two-target situation it
is possible that both targets are masked by each other. Weak targets
in the neighbourhood of strong targets are masked in almost all cases,
which reduces the range resolution and is not acceptable in real radar
applications. The ordered statistic OS-CFAR has in all these multi-
ple target situations a much better performance compared to CA and
CAGO-CFAR procedures.
OS-CFAR is not based on the assumption of homogeneous clutter
inside the reference window. Therefore the window length can be ex-
tended in the OS-CFAR case without any disadvantages. For compar-
ison, OS-CFAR with N = 24 will even outperform classical CA-CFAR
Waveform Design, Range CFAR and Target Recognition 25
with N = 16. This is an important advantage of the OS-CFAR proce-
dure.
The general idea of an ordered statistic is technically simple. To
estimate the average noise and clutter power a single rank X
(k)
of the
ordered statistic is used instead of the arithmetic mean. In this case
a very few large amplitudes in the sliding window have a very small
eect on the estimation result. OS-CFAR is robust in multiple target
situations. The threshold is hardly inuenced by a second or third target
inside the window.
All amplitudes inside the sliding window are sorted according to in-
creasing magnitude.
X
(1)
X
(2)
X
(N)
Z = X
(k)
. (33)
The pdf of the kth value of the ordered statistic is given by
P
X
(k)
(x) = p
k
(x) = k
_
N
_
(1 P
X
(x))
Nk
(P
X
(x))
k1
p
x
(x). (34)
Thus the pdf of the kth value of the ordered statistic for exponentially
distributed random variables is given by
P
X
(k)
(x) = p
k
(x) = k/
_
N
_
_
e
x/
_
Nk+1
(1 e
x/
)
k1
. (35)
The relation between P
fa
and T
OS
can be calculated by combining
(2) and (13)
P
fa
= k
_
N
k
_
(k 1)! (T
OS
+N k)!
(T
OS
+N)!
. (36)
The importance of this case is that OS-CFAR can be analytically
analysed without any approximations. Furthermore the resulting scal-
ing factor T
OS
is completely independent of . Figure 14 shows the
typical behaviour of OS-CFAR in clutter edge and multiple target situ-
ations. The threshold follows the clutter contour with a certain safety
distance. In two target situations the threshold is more or less unchanged
compared with a pure noise situation.
Nadav Levanon [21, 22] has applied OS-CFAR to a Weibull distributed
background signal and described the results analytically. Blake [26] anal-
ysed OS-CFAR in non-uniform clutter. Weber and Haykin [24] have
extended OS-CFAR to a two parameter distribution with variable skew-
ness.
26

Figure 14. OS-CFAR (N=24), k=3/4*N, P
fa
= 10
6
In [28] the performance of OS-CFAR in a 77GHz radar sensor for
car application is examined. In the automotive radar application case
multiple target situations occur almost always.
OSGO-CFAR. An extension of OS-CFAR has been developed by He
You in his paper [20]. In this case an ordered statistic is applied in the
left and right window part separately followed by a greatest of selection.
Therefore, the procedure is called OSGO CFAR. The computational
complexity is reduced in this case and the detection performance shown
in Figure 15 is good.
Gaspare Galati et. al. compared the performance of OS-CFAR and
OSGO CFAR in the presence of dierent backgrounds [23]. He found
that the OSGO method suers only a small additional loss with respect
to the OS. In a non-homogeneous background with clutter edges it even
shows superiority in the control of the false alarm probability.
Censored CFAR. Richard and Dillard [19] have proposed a CFAR
procedure which is based on CA-CFAR but is already close to the gen-
eral OS-CFAR idea when they are calculating the largest m values inside
the sliding window and excluding these values from the arithmetic mean
calculation. This step makes modied CA-CFAR less sensitive in mul-
tiple target situations. In their paper they analysed a Censored-CFAR
detector for m equals 1 and 2.
Waveform Design, Range CFAR and Target Recognition 27
Figure 15. OSGO-CFAR (N=24), k=3/4*N, P
fa
= 10
6
The resulting false alarm probability for m=1 can also be calculated
independently of the actual noise level:
P
fa
= (N1)
_
N
N 1
N1

k=1
_
N 1
k
_
(1)
k+1
(1 +k +T)
1
_
N1
. (37)
Figure 16 demonstrates the performance of a censored-CFAR detec-
tor with m=1. Compared with CA-CFAR the results here are better,
because in all scenarios the targets are detected. But compared with
OS-CFAR the threshold in the 2 target scenario is still too high in the
neighborhood of the targets. Ritcey [29] studied the performance of this
method for multiple target situations.
Figure 16. Censored-CFAR (N=24), k=1, P
fa
= 10
6
28
WCA-CFAR. If a-priori information about the target position is
made available by the tracking system the adaptive threshold can be
lowered. A CFAR method called weighted CA-CFAR uses this and was
proposed by Barkat, Himonas and Varshney [30].
This method separates the window cells into a leading and a lag-
ging part. Before the mean values of these parts are averaged, they are
weighted by the factors and . Optimum values for and are cal-
culated in accordance with the level of interference of present targets, so
that a constant false alarm rate and a high detection probability can be
guaranteed.
Other publications. Many additional papers have been published
which are based on the four fundamental procedures CA-, OS-, GO-
and SO-CFAR. In these publications either new combinations and mod-
ications of existing procedures or the performance in new environments
are analysed.
The main relevant models for radar detection are
Pure clutter or noise situations with fully homogeneous statistic
Transition between noise and clutter or between two clutter areas
with dierent average power
Non-homogeneous clutter
Correlated clutter
Dierent clutter amplitude statistics (Rayleigh, Weibull, . . . )
Dierent target amplitude statistics (Rice, Swerling models, . . . )
Single target situations
Two or multiple target situations.
A simple quantitative comparison between dierent CFAR procedures
in pure clutter and noise situations, which is the most important situ-
ation, can be calculated using the average detection threshold (ADT)
[14].
In this case the expectation of all calculated thresholds is calculated
and can be used for system comparison.
5. Conclusion
The objective of this chapter was to discuss some important contri-
butions for radar system design and digital radar signal processing. The
Waveform Design, Range CFAR and Target Recognition 29
focus was on waveform design in general and on automotive applica-
tions in particular. Target detection is an important issue for all radar
systems. Therefore some range CFAR procedures have been discussed
which can be applied especially in multiple target situations to avoid any
masking situations. Additionally some new results have been discussed
for target recognition.
References
[1] Artis, Jean-Paul; Henrio, Jean-Francois: Automotive Radar Development
Methodology, International Conference on Radar Systems, Brest, France, 1999.
[2] Klotz, Michael; Rohling, Hermann: A high range resolution radar system net-
work for parking aid applications, International Conference on Radar Systems,
Brest/ France, 1999.
[3] Klotz, Michael; Rohling, Hermann: 24 GHz Radar Sensors for Automotive Ap-
plications, International Conference on Microwaves and Radar, MIKON2000,
Wroc law/ Poland, 2000.
[4] Rohling, Hermann; Meinecke, Marc-Michael; Mende, Ralph: A 77 GHz Au-
tomotive Radar System for AICC Applications, International Conference on
Microwaves and Radar, MIKON98, Workshop, Krakow/ Poland, 1998.
[5] Rohling, Hermann; Meinecke, Marc-Michael; Klotz, Michael; Mende, Ralph: Ex-
periences with an Experimental Car controlled by a 77 GHz Radar Sensor, In-
ternational Radar Symposium, IRS98, M unchen, 1998.
[6] Stove, A. G.: Linear FMCW radar techniques, IEE Proceedings-F, Vol. 139,
No. 5, Oct. 1992.
[7] H. Rohling, A. Ho and U. L ubbert: Multistatic Radar Principles for Automo-
tive RadarNet Applications, IRS 2002 International Radar Symposium, Bonn,
Germany, 2002, pp. 181185.
[8] M. Schiementz, F. Folster and H. Rohling: Angle Estimation Techniques for
dierent 24GHz Radar Networks, IRS 2003 International Radar Symposium,
Dresden, Germany, 2003, pp. 405410.
[9] A. Hoess et al.: The RadarNet Project, 7th ITS World Congress, Torino,
November 2000.
[10] Brian Ricket: A Vision Of Future Applications For An Automotive Radar Net-
work, WIT 2004, 1st International Workshop on Intelligent Transportation,
Hamburg, Germany, 2004, pp. 117-121.
[11] N. Levanon: Radar Principles, Wiley & Sons, New York, 1988.
[12] D. Oprisan and H. Rohling: Tracking Systems for Automotive Radar Networks,
IEE Radar 2002, Edinburgh, UK.
[13] H. Rohling, F. Folster, F. Kruse and M. Ahrhold: Target Classication Based
on a 24GHz Radar Network, Radar 2004 Conference, Toulouse, France.
[14] H. Rohling: Radar CFAR Thresholding in Clutter and Multiple Target Situa-
tions, IEEE Trans. Aerosp. Electron. Syst. 19 No. 4, 1983, pp. 608-621.
[15] M. Klotz: An Automotive Short Range High Resolution Pulse Radar Network,
PhD thesis, TU Hamburg-Harburg, 2002.
30
[16] S. Blackman and R. Popoli: Design and Analysis of Modern Tracking Systems,
Artech House, Boston, 1999.
[17] Nitzberg R.: Analysis of the Arithmetic Mean CFAR Normalizer for Fluctuating
Targets, IEEE Transactions on AES, 1, pp. 44-47, 1978.
[18] Hansen, V.G., Sawyer, J.H.: Detectability loss due to greatest of selection in a
cell-averaging CFAR, IEEE Transactions on AES, 16, pp. 115-118, 1980.
[19] Richard, J.T., Dillard, G.M.: Adaptive Detection Algorithms for Multiple Target
Situations, IEEE Transactions on AES, 4, pp. 338-343, 1977.
[20] He You: Performance of Some Generalized Modied Order Statistics CFAR
Detectors with Automatic Censoring Technique in Multiple Target Situations,
IEE Proceedings F, 131(4), pp. 205-212, 1994.
[21] Levanon, N.: Detection loss due to interfering targets in ordered statistic CFAR,
IEEE Transactions on AES, (11/1988), pp. 678-681, 1988.
[22] Levanon, N., Shor, M.: Order statistic CFAR for Weibull background, IEE
Proceedings, F, 137, 3 (6/1990), pp. 157-162, 1990.
[23] Di Vito, A., Galati, G., Mura, R.: Analysis and comparison of two order statistics
CFAR systems, IEE Proceedings, F, 141, 2 (4/1994), pp. 109-115, 1994.
[24] Weber, P., Haykin, S.: Ordered statistic CFAR processing for two-parameter
distribution with variable skewness, IEEE Transactions on AES, AES-21, 6
(11/1985), pp. 819821, 1985.
[25] Weiss, M.: Analysis of some modied cell-averaging CFAR processors in
multiple-target situations, IEEE Transactions on AES, 18, pp. 102-114.
[26] Blake, S.: OS-CFAR theory for multiple targets and nonuniform clutter, IEEE
Transactions on AES, 24, 6 (11/1988), pp. 785-790, 1988.
[27] Gini, F., Farina, A., Greco, M.: Selected List of References on Radar Signal
Processing, IEEE Transactions on AES, 37, 1 (1/2001), pp. 329-359, 2001.
[28] Rohling, H., Mende, R.: OS CFAR performance in a 77GHz Radar Sensor for
Car Application, CIE International Conference on Radar, Peking, China, 1996.
[29] Ritcey, J. A.: Censored mean-level detector analysis, IEEE Transactions on
Aerospace and Electronic Systems, AES-22, 3 (July 1986), pp. 443-454, 1986.
[30] Barkat, M., Himonas, S.D., Varshney, P.K.: CFAR detection for multiple target
situations, IEE Proceedings, F, 136, 5 (10/1989), pp. 193-209, 1989.