Radar target classification using multiple

perspectives
M. Vespe, C.J. Baker and H.D. Griffiths

Abstract: The problem of radar target classification is examined for the case when more than one
perspective or viewing angle of the target is available to the sensor. Using full-scale target signature
measurements as the source data, it is shown how, for the first time, multiple perspectives enhance
the classification performance. Indeed this is the case even if only one additional perspective is
available for exploitation. Further, we explore the classification performance both as a function
of the number of perspectives and of the signal to noise ratio. Three approaches to high range
resolution profile multi-perspective classification have been implemented. This removes any poss-
ible bias that could be introduced by a single individual classifier. The results show, for all three, a
consistent improvement in the classification performance, as the number of perspectives is
increased. The techniques employed also provide considerable insight into the classification
process highlighting the degree of complexity of this extremely challenging problem.

1 Introduction networks of co-operating radar systems [7] offer a possible

solution for improving classification performance. These
The ability to detect and locate targets on a day/night, all have the added advantages of improved system sensitivity,
weather basis, over wide areas, has long made radar a key tailored coverage and reduced vulnerability to electronic
sensor in many military and civilian radar applications. It and physical attack.
is well recognised that the utility of the information supplied It is also worth noting that there are no rigorous and uni-
by a radar system would be hugely enhanced if targets could versally accepted definitions for the terms ‘classification’
additionally be classified. This might be by type (e.g. to and ‘recognition’. Here, The terms are used interchangeably
differentiate helicopters from civilian airliners or from mili- to address the ability of the radar system and subsequent
tary jets) or even better, within type (e.g. a Chieftain tank processing to determine the class to which a target
from a T32 tank). It would be even more advantageous if belongs (after defining a set of sub-populations such as a
this could be automated, or at least partially automated. In particular version of a tank). To perform automatic target
addition, huge quantities of imagery may be produced by recognition (ATR) in a useful manner it is necessary to
SAR systems and place unrealistic demands on a human achieve a sufficiently high degree of confidence that a
image interpreter. Algorithmic aids are necessary if the target has been correctly classified, and hence it is
data are ever to be fully exploited (or even reviewed). extremely important to minimise the probability of
However, it has proved far from straightforward to create misclassification. This is because in both military and civi-
robust, reliable classifiers capable of coping with wide lian applications there are likely to be safety of life aspects
target and clutter conditions. Most of the approaches to that demand an acceptably high level of performance whose
classification have concentrated on exploiting higher and limits and variability are fully understood.
higher spatial resolutions both in one [1] and two [2] dimen- When a radar system illuminates a target the energy is
sions. Research has also been reported examining the role of reflected in all directions with some of this signal returning
polarimetric information (e.g. [2, 3]) often in conjunction in the direction of the radar. The reflected signal may be
with high resolution. However, there has been little research modified because of a combination of the target itself, any
published that examines the utility of angular diversity for clutter in the illumination cell and often a forward scatter
improving classification performance [4– 6]. This is a component due to multipath. The classification system
little surprising as it is clear (1) there must be additional attempts to recognise the target based upon an analysis of
information on multiple perspectives and (2) it is extremely this reflected signal. It is clear that the contributions to the
easy for existing systems to acquire such data. A simple reflected signal from anything other than the target immedi-
example would be an aircraft flying past a scene of interest, ately begin to considerably complicate the problem. This
which will automatically generate multiple viewing angles. becomes even more severe as, because of the coherency
In fact, to date, the variation of target signature with of radar systems, there are fluctuations caused by any
angle has been a source of performance degradation in change in geometry and additionally, of course, the huge
single perspective classification approaches. Alternatively, range of possible targets that have to be distinguished
from one another. For these reasons, reliable target classifi-
# The Institution of Engineering and Technology 2007 cation has remained somewhat elusive.
doi:10.1049/iet-rsn:20060049 The ATR can be based either on 1D signatures, that is fea-
Paper first received 5th April 2006 and in revised form 8th March 2007 tures derived from a range profile of the target to be classified
The authors are with the Department of Electronic and Electrical Engineering, together with the pattern used to train and test the classifier or
University College London, Torrington Place, London WC1E 7JE, UK 2D imagery can be used. The former is often employed
E-mail: [email protected] because of its simplicity in terms of implementation and
300 IET Radar Sonar Navig., 2007, 1, (4), pp. 300 – 307
signal processing, and the fact that the profile is less vulner- Here, we briefly describe the radar system, dataset and
able to motion-induced corruption. However, the use of 1D sub-population problem. This includes the data processing
profiles in an automatic classifier can lead to performance and the method of principal component analysis (PCA)
that is prone to low-classification rates because of the subsequently employed to generate the input to the classi-
additional inclusion of a clutter component that is scenario- fier. We then describe the implementation of three different
dependent. The 2D imagery is less likely to contain clutter multi-perspective (M-P) classifiers. The data output by the
(although there will still be a significant multipath or PCA are used to generate the results of the M-P classifier
forward component) and hence would seem to promise a as a function of the number of perspectives used and as a
more detailed and authentic (but nevertheless still corrupted) function of the signal to noise ratio. Finally, the findings
representation of the target backscatter signature [8]. The 2D of this research are summarised and conclusions drawn.
imagery also requires more sophisticated signal processing
and if the data are collected over an extended period of time 2 High-resolution signatures and sub-population
(as in a frequency-stepped radar system) correction for irregu- problem
lar target movement is required. Such a correction can be far
from straightforward, especially as the detailed form of the The time domain backscattering properties of radar targets
motion corruption is usually unknown. can be used to form feature vectors for subsequent classifi-
Herein, we assume that the target has already been detected cation. These backscattering properties can take the form of
and tracked by the illuminating radar system and that the either HRRPs or inverse synthetic aperture radar (ISAR)
reflection signature to be recognised is a 1D high range resol- imagery. Both are well known methods of expressing the
ution (HRR) profile. The complete database used is collected backscatter from a radar target [9]. HRR data of full-scale
at all angles (i.e. over 3608) with the radar system stationary real vehicle targets from the Thales airborne data acqui-
and receiving the echoes from a target mounted on a rotating sition system (ADAS) has been used as the signature set
platform. The illumination patch includes both the turntable to examine M-P target classification. The data collection
and a small amount of clutter from scattering in the area just geometry is depicted in Fig. 1. Two stationary trihedral
in front of it. Here, we have minimised the effects of noise targets are located in front and behind a turntable and
clutter, forward scattering and external interference via provide a means for computation of the radar point spread
clutter subtraction and use of a multipath-free geometry. function. For collection of the signatures for classification
Although this is not wholly realistic for real-world cases, the two rotating corner reflectors are replaced by vehicles.
the objective here is to determine whether or not multiple In the figure two moving reflectors are located on opposite
viewing perspectives improve classification performance. sides of the rotating turntable. During the 3608 target
This of course implies that there is an in-built assumption rotation, HRR profiles are collected at an angular interval
that any improvement observed will also be apparent for of every 2 min of arc. High resolution is achieved by trans-
targets in more natural environments. However, because of mitting a linear frequency-modulated (LFM or chirp) pulsed
the additional variability that will occur this assumption will waveform [10]. On reception pulse compression is utilised
require careful testing and verification. to generate an HRRP. For example for a pulse duration tp ,
For HRR profiles (HRRP), slightly different target orien- the chirp waveform is given by
tations, because of rotational or translational motion of the !
target (assuming no range migration occurs) change the rela- 1 t
tive phase between scatterers whose returns go into the same p(t) ¼ pffiffiffiffi rect e jt(vc þpgt) (1)
tp tp
range bin and hence cause interference resulting in fluctu-
ations in the range profile. This is known as speckle. The
aspect angle is also a factor in determining shadowing where vc is the carrier frequency and g the rate of frequency
phenomena: a number of scatterers may be occluded by variation with time. The instantaneous frequency is
other parts of the target (local shadowing) or by other 1 d  v
targets (global shadowing) and therefore not be seen by the f (t) ¼ vc t þ pgt2 ¼ c þ gt (2)
2p dt 2p
incident waveform. As a consequence, a portion of the
target is unavoidably masked and the signature further Therefore, the bandwidth is B ¼ gtp . The pulse com-
corrupted. This may be thought of as a loss of information pression ratio (PCR) is equal to the time-bandwidth
as it can often result in significant regions of the target not product Btp and represents the range compression factor
contributing to the measured signature or profile. The (i.e. the ratio of the pulse’s length before time compression,
employment of multiple perspectives is a way to overcome tp , to its length after compression, 1/B). The PCR is a gauge
some of these effects through exploitation of the additional of the range resolution improvement given by pulse modu-
information that exhibits aspect angle dependency. Multiple lation. The ADAS data chirp parameters are shown in
perspective classification using 2D imagery is an additional Table 1. These enable HRRPs with a range resolution of
alternative but is not within the scope of this paper. just over 30 cm after weighting. The format of the data is

Fig. 1 ISAR geometry: two stationary corner reflectors are in front and behind the turntable, while two rotating ones are placed on the
turntable
IET Radar Sonar Navig., Vol. 1, No. 4, August 2007 301
such that a series of pulses can be collected where the band-
width of each pulse is stepped by half the bandwidth of any
given pulses. These can be combined to provide for very
high-range resolution signature with a resolution of 6 cm.
However, here we limit investigations of M-P classification
performance to the 30 cm range profiles only.
Fig. 2 shows an example of the form of the radar data. It
depicts the intensity modulated time history of HRRPs from
the two stationary and the two rotating trihedral reflectors
for the geometry depicted in Fig. 1. The two stationary tri-
hedrals show a constant response at near and far range as
expected. For the two rotating trihedral targets, when the
line-of-sight is on the trihedral bisector, a peak of reflection
occurs. This is consistent with the expected theoretical
response [11]. As the trihedral targets rotate, the backscat-
tered field decreases progressively until a point is reached
where there is a peak of specular reflection. This is a reflec-
tion from one of the sides making up the trihedral which is Fig. 2 History of HRR range profiles (30 cm range resolution)
orthogonal to the illuminating radar system (i.e. it faces the from four corner reflectors, two rotating and two stationary
radar beam and looks like a flat plate reflector). At increas- Zero degrees angle is in correspondence with the two corners having
ing rotation angles the RCS of the target drops because the the same range relative to the system and directly facing the radar
beam
orientation of the trihedral is such that it tends to reflect inci-
dent radiation away from the radar. This angular depen-
dency of the RCS of a well known reflector, such as a profiles spanning 3608. The data are divided into contiguous
trihedral, begins to illustrate how the backscattering proper- 108 sections, and 36 signatures randomly selected from each
ties of real targets may vary with the orientation of obser- of the 108 sectors are given to the classifier to build the
vation. For example if a target has part of its structure internal representation of the class. These signatures are
that mimics a trihedral it will only show this feature over then removed completely from the data set. The remaining
a limited angular range. Thus in a M-P environment, differ- profiles then provide the input to the classifier.
ent angular samples of a target signature should improve the
likelihood of observing a corner or corner-like reflector. 3 Data processing
Such shapes can be common on many manmade structures
and are often quite dominant features that may prove useful After pulse compression and weighting, the signal is Fourier
for classification. In Fig. 3, the complete angular ambit of transformed to form the total range profile. Subsequently,
range profiles spanning 3608 from a Land Rover vehicle the HRR profile is cut out of the illuminated area preserving
rotating on the turntable is shown. This highlights a 52 range bins, each with a sampled resolution of 30 cm, and
number of different scattering behaviours: the strong an actual resolution of approximately 39cm (because of the
peaks from specular reflections (08, 908, 1808, . . .) appear window applied). However, the data exhibit a low-level
over a very limited angle range and obscure nearby point- contribution from zero doppler clutter (ZDC). ZDC is
like backscattering. Corner-like returns can be observed at stationary or quasi-stationary clutter that contributes to the
a far range (6 m) for two range angular spans ([108 – HRRPs. By observing a sequence of range profiles it
608] and [1308 – 1808]). These returns correspond to the becomes clear how the ZDC affects the classification
trihedral-like structures formed at the rear of the Land process. Grass and trees show a relatively fast fluctuating
Rover. This is a vehicle without the rear soft top and has term whose effect at DC can be observed over relatively
a metallic bench seat that makes a corner where it joins small turntable rotation angles. This is in contrast to the
the rear bulkhead of the driver’s cabin. At 8 m range
there is a double bounce return corresponding to one of
the corners. This type of effect increases the information
that can be extracted which would be otherwise impossible
to reconstruct by a traditional single-perspective approach.
Classification has subsequently been performed on a sub-
population consisting of a total of three ground vehicles (A,
B, and C) exploiting multiple perspectives of HRRPs from
each of the vehicles. For each target, there are 10 500

Table 1: ADAS single chirp parameters

pulse length T 341 ns

pulse repetition interval PRI 2.048 ms
centre frequency fc 9.25 GHz
chirp rate g 1.46
1015
pulse compression ratio PCR 170.5
transmitted bandwidth Btx 500 MHz Fig. 3 History of HRR range profiles (8 cm range resolution)
from a series of X-Band stepped frequency chirps illuminating a
sampling frequency fs 500 MHz
ground vehicle as it rotates over 3608
range resolution Dr 30 cm At zero degrees, the target is broadside oriented, while at 908 it is
end-on to the radar

302 IET Radar Sonar Navig., Vol. 1, No. 4, August 2007

more persistent response given by spatially more restricted The test and training feature vectors can then be trans-
scatterers such as static ground discontinuities and mounds formed as follows
of gravel. This also means that the ZDC can be very differ-
ent from one target measurement to another because as the f 0 ¼ V T (fn  f
) (4)
vehicle approaches the turntable it modifies the local clutter
The number of principal components PC ¼ P is chosen as a
structure and hence the ZDC component of the return in the
function of the classification rate achieved. This usually
signature. As a consequence, if ATR is implemented using
becomes stable once the PCs necessary to fully describe the
the same data for both training and testing the classifier, the
data have been selected. After testing the classifiers, their
overall accuracy could be significantly upset by the ZDC
mean value correct classification rates (CCRs) are plotted
contribution to backscattering, as the classifier would other-
against the number of principal components representing the
wise use the clutter as an input. Thus the ZDC has been esti-
feature vectors. For the data used here the dimension of
mated and subtracted. This has been performed using data
the feature vector is 52. However, as can be seen in Fig. 4,
derived in the frequency domain following the method out-
the probability of correct classification PCC is almost
lined in [12]. The mean value of ZDC over about 508 of
unaltered until less than approximately 12 components. At
turntable rotation is subtracted from those frequency signa-
this point they are no longer representative of the different
tures, and then the final HRRP is extracted.
target classes. Therefore, in order to depict the target with
After ZDC subtraction, a mean filter over a neighbourhood
more robustness, the feature vectors used in the following sec-
of three echoes is applied to a sequence of range profiles.
tions are obtained using the first 21 principal components.
Since two consecutive HRRPs are spaced by only 2 of
target rotation the non-coherent averaging partially averages
out those phenomena that are independent of the geometry. 5 Multi-Perspective classifier implementation
In addition, the signal to noise ratio (SNR)
p is increased as
In order to remove any bias introduced by a single algorithm
demonstrated in [13] by a factor N where N is the
three classifiers have been implemented. These are: (1) a
number of target echoes. Prior to non-coherent averaging,
naïve Bayesian classifier, (2) a nearest neighbour classifier
it is usually necessary to align the range profiles. However,
and (3) a neural network classifier. All three are described
here, because the input patterns for recognition are collected
in detail below. A traditional single-perspective classifier,
from the turntable measurements, no additional alignment is
after training the ATR algorithm with a set of templates, is
needed between training and test profiles.
tested with profiles collected from all orientations of the
Finally, the signatures are normalised following [14] to
target. In an N-perspective scenario, the parameter that
make the target area in the HRRP more prominent. A
distinguishes the perspective topology is the vector F ¼
target mask is generated to select the target only as part
fDfi,j: i, j ¼ 1, . . . , Ng where the element Dfi,j ¼ fi 2 fj
of the range profile by measuring the noise mean value m
represents the angular displacement between the node i
and standard deviation s. A threshold m þ 3s is applied
and j. For example in a two-perspective (2-P) scenario, the
to eliminate the non-target zone. The target masks are
parameter that distinguishes the perspective node locations
used so that target-only data is presented to the classifier.
is their relative angular displacement Df1,2 ¼ f2 2 f1 .
Hence, after fixing Df1,2 , the 2-P classifier is tested with
4 Dimensional reduction: principal components all possible pairs of HRR profiles displaced by that angle
analysis covering all the possible orientations of the target. Having
a test set consisting of N profiles, the same number of pairs
In a typical pattern recognition problem, it is often necess- can be formed to test the 2-P classifier.
ary to reduce the data dimension of the input of the classi- As shown in Fig. 5, to test the M-P classifier, the
fier. This is mainly because of an intrinsic degree of sequence of all the possible pairs of HRR profiles displaced
redundancy within the data where there are regions where by Df1,2 is given as an input, and the accuracy for a 2-P
no information is contained. It also helps to reduce the classifier is subsequently computed. Thus to evaluate the
very high quantities of data that otherwise have to be pro-
cessed. Thus we may consider range profiles as providers
of feature vectors that are to be separated. After focusing
on those range resolution cells whose intensities depict
the target backscattering, the number of elements represent-
ing the backscatter can be reduced with an information loss
which is assumed negligible. Furthermore, dimensional
reduction of this type also attempts to emphasise the differ-
ences between patterns and hence enhance classification
performance.
The PCA [15] is a statistical method that enables the data
to be represented in a different vector basis such that it is
possible to remove similarities (which therefore do not con-
tribute to the classification process). After subtracting the
mean f̄ from each of the vectors of the training set F and
producing a zero-mean set of data, the covariance matrix
Cov can be formed
1X
Cov(F) ¼ (f  f
)  (fn  f
)T (3)
N n n Fig. 4 Classification accuracy against number of PCs used
The correct classification rates remain almost unchanged using more
After calculating the Eigenvectors of Cov(F ), the P-most than PC ¼ 13
significant eigenvectors with the largest Eigenvalues are Then the accuracy drops, signifying that the information content of the
selected and form a new basis vector V ¼ (v1 , v2 , . . . , vP). feature vector is not reliably represented in terms of classification

IET Radar Sonar Navig., Vol. 1, No. 4, August 2007 303

3-P recognition performance, after choosing the set to pure calculations of feature probabilities. The training
F ¼ fDF1,2 , DF1,3g, the classification accuracy is measured set is stored in different databases of templates and therefore
as the average correct classification over all the possible the number of classes is known as well as the sets of their
HRR profiles triples that could be formed from the M-P representative vectors (i.e. the learning strategy is super-
topology F. This procedure is extended for networks vised). We consider a set of nc classes fCi : i : 1, . . . , ncg,
having four and five perspectives. and a single HRR profile X formed by a sequence of n
The training set of representative vectors for each class is elements x1 , x2 ,. . . , xn as the feature vector. The classifier
made up of 36 range profiles, taken approximately for every decides that X belongs to the class Ci showing the highest
108 of target rotation. The ‘testing set’ of each class consists posterior probability P(CijX ). This is done using Bayes’
of the remaining range profiles neglecting the templates. The theorem to calculate the unknown posterior probability of
angular displacement between nodes is not processed as classes conditioned on the unknown feature vector to be
information by the M-P classifier. Although this is only classified
one of the possible approaches to M-P classification it is
simple to implement. Furthermore, the computational P(X jCi )P(Ci )
P(Ci jX ) ¼ (5)
burden is reduced to a minimum. Although the target has P(X )
already been assumed to have been detected and tracked,
the angular displacements F of the nodes are unknown to The naïve Bayesian classifier is based on the assumption
the network. Therefore, in case of a single sensor collecting of class independence of each attribute of the feature vector.
multiple perspectives from a moving target, it is assumed Furthermore, because the values x1 , x2 , . . . , xn are continu-
that to combine the signatures, it is not necessary to estimate ous values, they are assumed Gaussian distributed. Their
the motion parameters and the centre of rotation of the target. statistical parameters are deduced from the training set
Here, we consider a network of radars consisting of N and used to calculate the conditional probability P(xijCi)
nodes, each single perspective j of the target is represented for the attribute xi and, eventually, the likelihood ratio test
by the HRR profile Xj . Consequently, the set of signatures P(X jCi ) P(Cj )
collected by the network is represented by fXj : j ¼ 1, . . . , If . ) X [ class‘i’ (6)
Ng and is directly related to the vector of angular node P(X jCj ) P(Ci )
displacements F. In the next section the signatures fXj:
j ¼ 1, . . . , Ng correspond to the HRRPs after applying PCA. These concepts are integrated for the M-P naïve Bayesian
The following sections explain the specific M-P classifier. Here, we consider a network of N radars and the
implementations used for the three different approaches sequence of 1D signatures fXj : ¼ 1, . . . , Ng is the infor-
for classification. Fig. 6 illustrates the particular procedure mation collected by the system. The posterior probability
which consists of three separate stages: the first is a single- P(CijX1 , . . . , XN) of the sequence of range profiles con-
perspective stage, where the ith perspective is partially pro- ditioned to Ci is the probability that the sequence belongs
cessed to generate the feature vector. The second is the M-P to that class and by applying Bayes’ theorem, can be
stage, where the single-perspective outputs are processed expressed as follows
and combined. The third is the classification stage where YN
P(Xj jCi )P(Ci )
P(Ci jX1 , . . . , XN ) ¼ (7)
the final decision is made. This structure allows for parallel j¼1
P(Xj )
implementation of the procedure and therefore reduces the
Assuming the probability P(X1 , . . . , XN) with a fixed
time of execution.
number of perspectives to be constant, the final decision
is made for the class Ci that maximises P(CijX1 , . . . , XN).
5.1 Naïve Bayesian multi-perspective classifier This procedure enables the distinction of a single-
perspective stage where all the conditional probabilities
The naïve Bayesian classifier is a statistical method of P(XjjCj) are computed separately for each perspective.
pattern recognition [16]. The decision-making is reduced
5.2 K-NN Multi-perspective classifier

The K-nearest neighbours (K-NN) is a decision-theoretic

approach to classification [17]. The rule consists of measur-
ing and selecting the minimum K distances from the feature
vector to be classified and comparing with the templates of
the different classes. Consider an input vector X and a popu-
lation made up of a set of classes fCi : i ¼ 1, . . . , ncg, then

Fig. 5 Multi-perspective environment approximation

Considering a 2-P problem, for all target orientations c covering 3608,
Fig. 6 Multi-perspective classifier consists of a first single-
the classifier is tested with all possible pairs of HRR profiles displaced perspective stage that enables the parallelisation of the process
by the angle D1,2 nodes independently produce those partial outputs that are sub-
This is then extended up to five perspectives and classification rates sequently processed by the M-P stage in order to eventually classify
computed the object

304 IET Radar Sonar Navig., Vol. 1, No. 4, August 2007

the distances dt,i from X to Tt,i , the tth template vector of the vector X and a population made up of a set fCi :
ith class, can be computed and stored i ¼ 1, . . . , ncg of nc classes, the execution mode consists
of the calculation of the output vector Y ¼ ( y1 , y2 , . . . ,
dt,i ¼ d(X ,Tt,i ) ¼ kX  Tt,i k (8) ync). Since the unipolar sigmoid (or logistic) was chosen
as the activation function, the elements of the output
Subsequently, the K minimum scalars dt,i are selected vectors range from 0 to 1. The ultimate decision is made
from each class forming a K-dimensional vector D labelled for the ith class, where i is the index of the maximum
in ascending order. The final decision is made on the basis value of Y.
of the largest number of votes over the K dimensional vector The M-P FANNs classifier is structured following the
obtained. M-P scheme depicted in Fig. 6. If we assume N perspec-
There are three stages of the M-P K-NN classifier which tives, the sequence of feature vectors fXj : ¼ 1, . . . , Ng is
are implemented as follows: the input for the first stage. Each single-perspective
network accepts a vector and the partial outputs
1. The mono-perspective stage: after the collection of the
sequence of feature vectors {Xj : j ¼ 1, . . . , N} where N is fYj: ¼ 1, . . . , Ng are calculated. Subsequently, the output
the number of sensors in the network, the same number of Ȳ of the M-P stage is the mean value of the partial outcomes,
single-perspective classifiers is implemented. The jth classi- and the classification decision is finally made on the basis of
fier computes a vector Dj consisting of the K minimum dis- the maximum index of Ȳ.
tances from the templates.
2. M-P processing: the whole of the vector Dj is processed, 6 Multi-perspective classification performance
and the minimum K distances are selected giving a weight
for the decision. Classification performance is evaluated in a forced
3. Classification: the sequence of feature vectors in the decision environment using confusion matrices (e.g. see
input is associated with the class with the greatest number [18]) where each column represents the instances in a pre-
of weights. dicted class, whereas the rows represent the output distri-
bution for an actual class. In Table 2, the confusion
Different K values have been tested for this problem: the matrices of the three classifiers are shown with respect
best trade-off between complexity and correct classification to the number of perspectives involved in the decision
rates suggests a value K ¼ 5 for minimum distances. processing. The single-perspective results represent the
more usual single perspective classifier performance. It
5.3 FANNs Multi-perspective classifier should be noted that even with a single perspective the
classification performance is very good. This is a
Given a feature vector X, artificial neural networks (ANNs) function of using turntable data together with a small
learn how to execute the classification task by means of number of sub-populations. However, it does allow the
examples. They are able to analyse and associate an trends with a number of perspectives to be explored.
output corresponding to a particular class of objects [17]. Table 2 clearly shows an improvement in classification
In this paper, feed-forward ANNs (FANNs) supervised performance, as the number of perspectives is increased.
with a back-propagation strategy have been implemented. This trend is observed for all three classifiers and
During the learning phase the training samples are used to hence we can conclude that it is not a function of the
set internal parameters of the network, that is, after giving classification process but because of the additional
the templates as inputs to the classifier, the weights are perspectives. Interestingly, the biggest increase in
modified on the basis of the distance between the desired performance occurs when just one additional perspective
and actual outputs of the network. Considering an input is employed.

Table 2: Multi-perspective confusion matrices for three classifiers (K-NN, naïve Bayesian and FANNs) using different
numbers of perspectives

Classifier K-NN, % Bayesian, % FANNs, %

persp.: class A B C A B C A B C

1 A 79.37 1.72 18.91 78.39 6.90 14.71 89.73 1.4 8.87

B 2.92 85.82 11.26 2.10 67.99 29.91 2.58 82.06 15.36
C 1.92 20.85 77.23 4.86 10.34 84.80 1.63 11.49 86.88
2 A 91.61 0.48 7.91 90.38 1.74 7.88 96.77 0.15 3.08
B 1.11 94.92 3.97 2.04 77.82 20.12 0.39 94.58 5.03
C 0.25 15.97 83.78 3.83 8.20 87.97 0.45 5.97 93.58
3 A 92.26 0.07 7.67 95.27 0.71 4.02 98.12 0.05 1.83
B 0 97.88 2.12 1.84 84.52 13.64 0.19 96.41 3.40
C 0.04 7.59 92.37 2.58 6.68 90.74 0 4.03 95.97
4 A 96.88 0.01 3.11 98.24 0.19 1.57 99.44 0 0.56
B 0 99.65 0.35 1.55 89.86 8.59 0.05 98.66 1.29
C 0 7.05 92.95 1.78 5.66 92.56 0 2.14 97.86
5 A 97.34 0 2.66 98.93 0.09 0.98 99.75 0 0.25
B 0 99.82 0.18 0.94 95.18 3.88 0 99.36 0.64
C 0 3.5 96.5 1.09 4.99 93.92 0 1.20 98.80

IET Radar Sonar Navig., Vol. 1, No. 4, August 2007 305

 A parameter to measure the goodness of a classifier is
the CCR (i.e. the Probability of Correct Classification
PCC on a finite test dataset). It is the mean value of the
elements on the diagonal of the confusion matrix. In
Table 3, the CCR of the three implemented classifiers
are displayed using different numbers of perspectives.
Again all three show the same trend of improvement in
performance as the number of perspectives is increased.
Fig. 7 shows this graphically and again highlights that
the increase in classification performance is independent
of the particular classifier approach used. Furthermore, it
emphasises that the M-P improvements are not linear
with the number of perspectives involved. On average
an 8.8% improvement in the correct classification rate,
with respect to the single perspective classifier, is
reached when two perspectives are used, whereas the
benefit in relation to the 2-P classifier is 3.6% when Fig. 7 Correct classification rates of three classifiers using
three perspectives are used. It is 2.51% when going different numbers of perspectives
from three to four perspectives and finally 1.5% when
going from four to five perspectives. This is a conse-
quence of those signatures that relate to particular target range profile history and the SNRclass2i
orientations where specular reflections occur and the  
profile is thus dominated by only a very few scattering 1 XN X K
r
SNRclassi ¼ 20 log n,k (9)
centres. These patterns are common to all the target NK n¼1 k¼1 mn
classes and consequently are more easily misclassified.
Nevertheless, for these target orientations, the decision where mn is the mean noise intensity for the nth,
confidence of the classifier is low and is therefore signifi- k-dimensional profile RPn ¼ (rn,1 , rn,2 , . . . , rn,K), and is N
cantly aided by a second perspective. On the other hand, the number of profiles collected for the ith target class.
if both the first and second perspectives correspond to The result of applying thisP procedure to the original data
those with reduced information signatures, a third perspec- leads to an SNR ¼ (1/nc) iSNRclass2i ¼ 28.24 dB. The
tive is less effective in improving performance. A third profiles are subsequently progressively corrupted with
perspective has less influence to correct the weights of different levels of Gaussian noise applied to the I and Q
the others and thus the improvement from subsequent per- channels, before matched filtering. Then, the SNR is
spectives is progressively attenuated. This trend may be measured as above, and classification attempted on PCs
different if the perspective displacement of the nodes of extracted from the 1D signatures. The original training
the network is processed by the classifier as useful infor- dataset is preserved. As there was no bias introduced by
mation. Nonetheless, the nonlinear relationship between any of the three classifiers only the FANNs classifier is
CCR and number of radar perspectives suggests that used to examine performance as a function of SNR. In
only the employment of a small number of nodes in the Fig. 8, an example of a single range profile is plotted
network is necessary in order to achieve the best at different SNR levels. The radar length of the target at
trade-off between system complexity and improvement this heading is approximately 5.5 m, spanning the 17th to
in classification performance. the 35th range bins. As can be observed, the low-level
scatterers of the target gradually disappear in noise. In
addition the peak values become more and more corrupted
6.1 Classification performances against SNR
with the result of reducing the useful information in terms of
The benefits of using multiple perspectives to improve the higher peak values and their relative locations.
classification performance are now examined as a function Furthermore, the radar length of the target is no longer
of different SNR levels. The SNR of the data is measured deducible with any great certainty. Clearly this will have
as the distance in dB from the maximum magnitude in the an adverse effect on classification performance, no matter
HRR profile belonging to a particular target and the noise how many perspectives are used.
level after ZDC. From each target class ‘i’, the noise level In Fig. 9, the CCR results for just the FANNs classifier
is measured in the non-target area before target normalisa- are depicted against the number of perspectives used and
tion. The target signal level is estimated as the mean the SNR levels. In all cases increasing the number of
value of the maximum target backscattering over the 3608 perspectives improves classification performance. In
comparison to a single-perspective classifier, the CCRs
remain stable above a SNR threshold of 17 dB and then a
Table 3: Multi-perspective correct classification rates very rapid deterioration of the classification performance
(CCR) for the three classifiers using a progressive can be observed. This is because of the information loss
number of perspectives caused by the increasing noise that corrupts the range
profile. The relatively high SNRs at which this occurs can
K-NN, % Bayesian, % FANNs, % be partly explained by the fact that this is a combination
of smaller magnitude and larger magnitude scatterers that
1 80.81 77.06 86.22
provide the classification information. The smaller magni-
2 90.10 85.39 94.97 tude scatterers are more severely affected by the increasing
3 94.17 90.17 96.83 noise and hence quickly have an impact on reducing
4 96.49 93.55 98.65 performance. As the number of network nodes increases,
5 97.88 96.01 99.30 the information is partially recovered with the help of
other perspectives because the noise is not correlated and
306 IET Radar Sonar Navig., Vol. 1, No. 4, August 2007
 radar that obtains multiple perspectives as it traverses an
object or area of interest (provided sufficient angular
diversity is obtained). The performance has also been
investigated as a function of different noise levels. The
incidence of multiple perspectives on correct classification
performance improvement is observable for any value of
the SNR range of values examined. In addition, the M-P
benefits are more pronounced for low SNR levels that are
for those SNR values that are typical of not controlled
real-world experiments. Further research is required to
investigate the effects of the separation of perspectives
used to evaluate any relationship between angles of perspec-
tives selected and the resulting classification performance.
This is particularly important if single monostatic radar
systems are to be used as it could minimize the time at
which targets are illuminated.

8 Acknowledgments

Fig. 8 Range profile belonging to class A progressively cor- The work reported in this paper was funded by the
rupted with Gaussian noise on the I and Q channels Electro-Magnetic Remote Sensing (EMRS) Defence
Technology Centre, established by the UK Ministry of
Defence and run by a consortium of Selex, Thales
Defence, Roke Manor Research and Filtronic. The authors
would like to thank Thales Defence for providing ADAS
files to investigate HRR profiles from real targets.

9 References
1 Rong, H., and Zhaoda, Z.: ‘Researches on radar target classification
based on high resolution range profiles’. IEEE Aerospace and
Electronics Conf., July 1997, vol. 2, pp. 951– 955
2 Novak, L.M., Halversen, S.D., Owirka, G., and Hiett, M.: ‘Effects of
polarization and resolution on SAR ATR’, IEEE Trans. Aerosp.
Electron. Syst., 1997, 33, (1), pp. 102–116
3 Sadjadi, F.: ‘Improved target classification using optimum
polarimetric SAR signatures’, IEEE Trans. Aerosp. Electron. Syst.,
2002, 38, (1), pp. 38–49
4 Shihao, J., Xuejun, L., and Carin, L.: ‘Adaptive multiaspect target
classification and detection with hidden Markov models’, IEEE
Fig. 9 Correct classification rates for an M-P FANNs classifier Sens. J., 2005, 5, (5), pp. 1035– 1042
5 Xuejun, L., Runkle, P., and Carin, L.: ‘Identification of ground targets
with different SNR test sets
from sequential high-range-resolution radar signatures’, IEEE Trans.
Aerosp. Electron. Syst., 2002, 38, (4), pp. 1230–1242
6 Runkle, P.R., Bharadwaj, P.K., Couchman, L., and Carin, L.: ‘Hidden
hence modifies the profiles differently. This result does Markov models for multiaspect target classification’, IEEE Trans.
highlight the importance of having adequate SNR such Signal Process., 1999, 47, (7), pp. 2035– 2040
7 Baker, C.J., and Hume, A.L.: ‘Netted Radar Sensing’, IEEE Aerosp.
that small magnitude scatterers are preserved as an input Electron. Syst. Mag., 2003, 18, (2), pp. 3 –6
to the classification process. 8 Novak, L.M.: ‘A comparison of 1-D and 2-D algorithms for radar
target classification’. IEEE Int. Conf. on Systems Engineering,
7 Conclusions and summary August 1991, pp. 6 –12
9 Wehner, D.R.: ‘High resolution radar’ (Artech House, USA, 1995)
10 Sullivan, R.J.: ‘Radar foundations for imaging and advanced
Here, A novel radar target classification approach has been concepts’ (SciTech Publishing, 2004)
introduced. The method is based on the processing of real, 11 Knott, E.F., Shaeffer, J.F., and Tuley, M.T.: ‘Radar cross section’
full-scale target signatures collected at different perspec- (Artech House, 1985)
tives in a network of monostatic radars. In a first single- 12 Showman, G.A., Richards, M.A., and Sangston, K.J.: ‘Comparison of
two algorithms for correcting zero-Doppler clutter in turntable ISAR
perspective stage, depending on the methodology adopted, imagery’. Conf. on Signals, Systems & Computers, 1998, vol. I,
the 1D imageries are processed in parallel, providing the pp. 411– 415
following M-P stage with partial outputs eventually used 13 Skolnik, M.I.: ‘Introduction to radar Systems’ (McGraw-Hill, 1980)
for the final decision. The benefits of the M-P classifier 14 Zyweck, A., and Bogner, R.E.: ‘Radar target classification of
commercial aircraft’, IEEE Trans. Aerosp. Electron. Syst., 1996, 32,
implementation have been analysed and show a nonlinear (2), pp. 598– 606
but very clear CCR improvement with the number of 15 Theodoridis, S., and Outroumbas, K.: ‘Pattern recognition’ (Academic
perspectives. As the number of radar nodes increases, the Press, 1999)
M-P CCR stabilises, suggesting that most benefit accrues 16 Looney, C.G.: ‘Pattern recognition using neural networks’ (Oxford
when using a small network of only two or three nodes University Press, 1998)
17 Duda, R.O., Hart, P.E., and Stork, D.G.: ‘Pattern classification’ (John
and hence avoids any significant increase of system Wiley and Sons, 2001, 2nd edn.)
implementation complexity. For example, this enables a 18 Tait, P.: ‘An introduction to radar target recognition’ (IEE
simple approach to be adopted for a single monostatic publications, 2005), ISBN: 0863415016

IET Radar Sonar Navig., Vol. 1, No. 4, August 2007 307