Sei sulla pagina 1di 12

www.ietdl.

org

Published in IET Signal Processing


Received on 1st June 2007
Revised on 2nd November 2007
doi: 10.1049/iet-spr:20070082

In Special Issue on Signal Processing Techniques for ISAR and


Feature Extraction

ISSN 1751-9675

Recent developments in detection, imaging


and classification for airborne maritime
surveillance
N. Bon1 G. Hajduch2 A. Khenchaf 3 R. Garello4
J.-M. Quellec1
1
Thales Airborne Systems, Brest, France
2
CLS/Radar Applications Division (formerly BOOST Technologies), Brest, France
3
Laboratoire E3I2 (EA 3876)-ENSIETA, Brest, France
4
GET-ENST Bretagne, Brest, France
E-mail: nicolas.bon@fr.thalesgroup.com

Abstract: The role of maritime patrol missions is to monitor large oceanic areas. The authors focus on a complete
signal-processing sequence from the primary detection of the targets to their classification or recognition from an
airborne radar. Contrary to the classical approach in which detection, tracking, imaging and classification are
considered separately, here the authors propose an integrated strategy based on a close cooperation among
all of them. Recent developments in high range resolution target detection are presented and their
integration in the complete system is discussed to limit false alarms. High-resolution ISAR imaging of ships is
then tackled and associated with a feature extraction process and a support vector machine classifier. A set of
real data is used to illustrate the imaging and classification results.

1 Introduction of some basic cinematic information on pre-detected targets


including heading and speed, by integrating information
Maritime patrol radars are usually aimed at detecting and collected over successive antenna scans. The imaging
locating targets to provide global information on the tactical function is usually based on synthetic aperture radar (SAR)
situation of a geographical area. Most of them operate with a or inverse synthetic aperture radar (ISAR). The classification
scanning antenna and cover distances extending to the function begins with segmentation of the ship profile from
horizon. Mere detection and localisation of the target is the clutter background, so as to extract relevant features and
insufficient when the target class is unknown. Indeed, in a compare it with a target database.
wartime context, the identification as friend or foe is essential,
whereas in peacetime ship classification is essential for fishing This strategy of four independent steps suffers from the lack
monitoring, search and rescue, and detection of containers or of interactions between steps preventing the incorporation in
other lost objects. The class of the ship may also provide one step of prior information collected in other steps. One
useful information. Thus, the automatic target recognition is illustration of this limitation is extreme detection conditions,
undeniably a major advantage provided by the radar. for instance small boats or low radar cross section (RCS)
targets embedded in high sea state, where classical detectors
Classical systems are based on a series of four main show their limits. To circumvent this fact, recent works
functions applied sequentially and independently: detection, [2, 3] have reported high range resolution (HRR) detectors
tracking, imaging and classification. Detection over the based on the generalised likelihood ratio test (GLRT)
scanned area is basically performed on low or medium range and we proposed in [4, 5] to extend these detectors to range
resolution data [1]. The tracking function allows estimation and Doppler spread targets. Those detectors require

192 IET Signal Process., 2008, Vol. 2, No. 3, pp. 192– 203
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-spr:20070082

Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 22,2010 at 02:37:52 EDT from IEEE Xplore. Restrictions apply.
www.ietdl.org

incorporation of prior information on target size or on cinematic gravity waves and the speckle s is a rapid fluctuating
information which can be easily estimated by tracking. component corresponding to the capillary waves [7].

This paper will present performance enhancements of a The derivation of a detector under this model may be
complete integrated system by applying cooperation based on different strategies. One strategy involves Conte’s
between functions and will be organised as follows. We will asymptotic normalized match filter (ANMF) [8]. His
first present in Section 2 how the detection and tracking detector is derived asymptotically assuming prior texture
functions contribute in a complete system for airborne distribution. In Gini [9] the texture is modelled as a
maritime surveillance. Each function will be discussed as deterministic unknown parameter estimated using the max-
well as the integration of recent developments in HRR likelihood method. It is interesting to note that both
detectors proposed in [4, 5]. More specifically, an extension approaches give the same test
of the range-spread detector to estimate the ship length
 
and Doppler frequencies will be presented and illustrated  H 1 2 H1
e M z 
by simulation results. Section 3 will be devoted to ship ,h
. (2)
target ISAR imaging and SVM automatic ship recognition. (zH M 1 z)(eH M 1 e) H0
The feature extraction process will be detailed and a
discriminate analysis of different features will be shown. The properties of such detectors have been widely studied,
The algorithms detailed in this section have been applied see for example [9, 10]. The threshold h is set to ensure a
on real data and results will be shown. constant false alarm rate (CFAR) property. A threshold set
too high entails the non-detection of small or low RCS
targets or targets embedded in high sea state. Another issue
2 Distributed target detectors for to address is the range resolution allowed by the radar
maritime surveillance radar systems waveform which generally consists in finding the best
trade-off between collapsing losses and high clutter-to-
A complete airborne radar system for surveillance missions is signal ratio in each range cell [5].
generally composed of four main functions: detection,
tracking, imaging and classification. They take place one
after another along the antenna scans, beginning with the 2.1 Distributed detectors
detection. Radar operators are very sensitive to false alarms To overcome these drawbacks, we propose to use HRR
and consequently the detection strategy is commonly based detectors such as reported [2, 3] or our generalisation to
on the Neyman– Pearson criterion to ensure a limited false range- and Doppler-spread targets in non-Gaussian clutter
alarm rate. The detection problem of ship target can be described in [4, 5]. These detectors can be seen as 1D or
formulated as binary hypotheses testing problem 2D matched filters to the target response. The previously
mentioned compromise on range resolution then
H0 :z ¼ c disappears. A key advantage of these detectors depends
(1)
H1 :z ¼ x þ c upon the fact that increasing the range resolution of the
radar reduces the clutter energy in each range cell. In
The H0 hypothesis corresponds to the only presence of clutter addition, resolved scatterers are less fluctuating than a point
and the H1 hypothesis to the presence of clutter and target. A target fully contained in one range cell. Indeed, in low
clutter-dominant scenario is here addressed so that thermal range resolution, the target response is the vectorial sum of
noise is neglected. This situation occurs for example at its different scatterer contributions which entails great
short-ranges. The vector length is the number of pulses N fluctuations. Such targets are often represented as proposed
received while the antenna beam sees one target; z is the in Swerling [11] or Nakagami [12].
observation vector; x is the vector corresponding to the
backscattered target signal and the vector c corresponds to The target is assumed to be spatially distributed over L
the sea clutter. In the case of coherent detectors, the target range cells
vector is modelled as x ¼ ae, where e ¼ (1, e j2pfD , . . . ,
e j2pfD (N 1) )t is the steering vector corresponding to the H0 :zr ¼ c r , r ¼ 1...L
(3)
normalised Doppler frequency fD . H1 :zr ¼ xr þ cr , r ¼ 1, . . . , L

In the case of an HRR radar or in high sea states, clutter where zr ¼ (zr (0), zr (1), . . . , zr (N  1))t . The observations
statistics cannot be modeled as a Gaussian random process are assumed to be independent between each range cell. The
due to the observation of spikes; instead, they are usually sea clutter vectors are assumed to follow the same model as
pffiffiffiffi
modelled as a compound Gaussian vector and, more before, in all range cells, so that cr ¼ tr sr with
precisely, as a spherically
pffiffiffi invariant random vector (SIRV) sr ¼ CN (0, M), r ¼ 1, . . . , L. In the case of range-
[6] so that c ¼ ts , with s ¼ CN (0, M). This model distributed targets of detectors such as [2, 3], the target
takes into account the compound nature of the sea: the vectors are modelled as xr ¼ ar e, r ¼ 1, . . . , L and our
texture t is a slow fluctuating component that models the generalisation in [4, 5] to range and Doppler-spread targets

IET Signal Process., 2008, Vol. 2, No. 3, pp. 192 – 203 193
doi: 10.1049/iet-spr:20070082 & The Institution of Engineering and Technology 2008
Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 22,2010 at 02:37:52 EDT from IEEE Xplore. Restrictions apply.
www.ietdl.org

models different scatterers’ Doppler frequencies. The signal A priori information about the target length cannot be used
vector xr ¼ (xr (0), xr (1), . . . , xr (N  1))t in each range cell directly as an input to the distributed detector. Indeed, as
is the sum of the contribution of pr scatterers so that seen by the radar, this length varies with respect to the
cosine of the aspect angle. We then resort to the tracking
pr
X function to estimate target heading and consequently its
 
xr (n) ¼ ar,k exp j2pfr,k n , n ¼ 0, . . . , N  1 (4) aspect angle. This task is achieved using Kalman filtering
k¼1 methods [1, 13]. Fig. 1 presents the three-step system we
propose. The first step involves pre-detecting searched
Thus, the signal can be expressed with matrix formulation as targets. This is done by applying usual CFAR detectors
xr ¼ E r ar , where ar ¼ (ar,1 , ar,2 , . . . , ar,pr )t is the vector of such as cell-averaging CFAR [14], the threshold being
the scatterers’ complex amplitudes and the steering matrix assessed with a sufficiently high false alarm probability so
is E r ¼ {e2pjfr,k n }k[[1,...,pr ],n[[1,...,N ] . The range and Doppler- that hard-to-detect targets are not discarded. This pre-
spread target detectors are expressed as [4, 5] detection step enables us to limit the computational burden
by applying distributed target detection only on interesting
! zones where targets are potentially present. Next, resulting
X
L
zH 1 H1
r M zr 0 plots are tracked over a minimum of two antenna scans so
N ln H 1 ,h
. (5)
z r (M  Q )z
r r H
that aspect angle of each target is determined. The
r¼1 0
distributed target detector is then applied locally in a
sliding window, with a threshold that corresponds to a low
where Qr ¼ M 1 E r (E H 1 1 H
r M Er ) Er M
1
is the orthogonal false alarm probability. This strategy achieves superior
projector on the whitened subspace spanned by the columns of performance as compared with point-like target detectors
E r . We have shown that our range- and Doppler-distributed with the same false alarm probability.
target detector allows performance enhancement with
respect to point-like and range-only distributed target We stress the fact that prior knowledge about the target
detectors. It is also interesting to note that the case r ¼ 1 length does not mean necessarily knowing its exact value.
and steering matrix reducing to a steering vector gives a Consequently, distributed target detectors could be applied
GLRT identical to that given in (2). Moreover, considering immediately without preprocessing. As shown in [5],
a range-only distributed target, i.e. only with a steering detection probability decreases when the detector is not
vector per range cell, we find an equivalent expression to exactly matched to the target length. However,
Gerlach’s detector [2]. All these distributed target detectors performance still remains quite good and better than point-
and their properties (influence of clutter covariance matrix like target detection in low range resolution. The proposed
estimation, clutter-to-thermal-noise ratio, mismatch on approach can be used in a selective way so that detection
target length etc.) have been described in [5]. performance is improved for a specific class of target size
(e.g. small, medium or large surface vessels).

2.2 System integration Prior knowledge about the position of Doppler frequencies
From a more practical point of view, one may wonder how to still remains a difficult issue. Indeed, even if the target
integrate such detectors into a complete system. Indeed, they structure is known from a photograph or constructor plans,
all are based upon the assumption that the target Doppler the positions of scatterers in the range-Doppler plane
frequencies are known. Two different approaches are cannot be deduced easily. Moreover, the Doppler extension
possible. The first, which requires prior knowledge on the of the target depends on time as well as pitch and roll
target, is described in this section. The second, which motion amplitudes. Possible solutions may involve using
requires no information on the target, will be developed in precomputed responses from a 3D CAD models of the
the next section. target, using complex and computationally expensive

Figure 1 Integration of range-distributed target detection


Prior knowledge about target lengths enables fine segmentation from the sea clutter after the tracking function

194 IET Signal Process., 2008, Vol. 2, No. 3, pp. 192– 203
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-spr:20070082

Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 22,2010 at 02:37:52 EDT from IEEE Xplore. Restrictions apply.
www.ietdl.org

electromagnetic scattering methods based on finite difference logarithms of elementary SCR in each range cell
time domain (FDTD) methods or high-frequency
!
approximation and geometrical theory of diffraction (GTD) Lþr
X0 zr Qr zH
r
such as Xpatch [15]. j(Z, r0 , L) ¼ ln (8)
r¼1þr0 zr (M 1  Qr )zH
r

2.3 Target parameters estimation


Fig. 2 shows Monte-Carlo simulation results. The target is
This section develops the second approach in which prior assumed to be distributed over L ¼ 1, L ¼ 2 or L ¼ 4
knowledge about target parameters is not available and range bins. The target subspace is single dimensioned so
must be estimated. From a more practical point of view, that the steering matrices E r , r ¼ 1, . . . , L, are steering
estimating target parameters adds a non-negligible vectors er ¼ e, r ¼ 1, . . . , L, assumed known and equal.
computational burden. The previously mentioned pre- The detector of (5) is then Gerlach’s one [2]
detection strategy can thus be applied to constrain the
detection to interesting areas. !
X
L
jeH M 1 zr j2 H1
N ln 1   H 1  H 1  . h0 (9)
zr M zr e M e H ,
2.3.1 Target length estimation: We first deal with the r¼1 0
target length estimation. For this purpose, we consider a
sliding window in which target position r0 and length L The local value of the texture tr , r ¼ 1, . . . , L is assumed to
are estimated. ML-estimates are obtained by maximising follow a gamma distribution with a shape parameter
the loglikelihood function lH 1 under H1 hypothesis n ¼ 0:5. The clutter is assumed white so that M ¼ I .
Detection probability is here defined as the probability that
^ 1, L}
{r0 þ ^ ¼ arg max lH 1 (Z, r0 , L) (6) the test in (9) corresponds to H1 and at least one range bin
r0 ,L of target-estimated positions belongs to the target. As we
can see in Fig. 2a, detection probability increases with target
where Z ¼ [zr0 , . . . , zr0 þL ]. It is straightforward to show length L. Fig. 2b shows root mean square error (RMSE) of
that lH 1 is expressed as target length L and position r0 estimates.
Lþr
!
X0 N
lH 1 (Z, r0 , L) / ln 1
(7) 2.3.2 Target subspace estimation: The estimation of
r¼1þr0 zr (M  Qr )zH
r the target subspace consists of determining the number of
scatterers pr , r ¼ 1, . . . , L, in each range cell and their
The denominators of the sum of logarithms correspond to Doppler frequencies fr,k , r ¼ 1, . . . , L, k ¼ 1, . . . , pr . This
the energy of the received signal projected on the sea estimation thus yields imaging the ship targets in
clutter whitened subspace. The energy of sea clutter conjunction with the detection function. In order to
contained in each range cell is obviously non-zero, so the estimate signal subspaces dimension, we resort to
maximisation of (7) will maximise L. We then maximise a Rissanen’s MDL [16] criterion which is a function of the
function of the signal-to-clutter ratio (SCR) j. Indeed, this ratio of geometric and arithmetic means of the covariance
quantity will reach its maximum when the entire target will matrix eigenvalues. As we proposed in [5], superresolution
be just contained in the window. We then use the GLRT spectral estimation methods can be implemented such as
previously derived as it is expressed as the sum of rootMUSIC (polynomial version of MUSIC [17]) or

Figure 2 Monte-Carlo simulation results


a Detection probability of range distributed targets for L ¼ 1 (dash-dotted line), L ¼ 2 (dashed line) and L ¼ 4 (solid line)
b Root mean square error of length estimates L and target position r0 (lines with square markers) for for L ¼ 1 (dash-dotted line), L ¼ 2
(dashed line) and L ¼ 4 (solid line)

IET Signal Process., 2008, Vol. 2, No. 3, pp. 192 – 203 195
doi: 10.1049/iet-spr:20070082 & The Institution of Engineering and Technology 2008
Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 22,2010 at 02:37:52 EDT from IEEE Xplore. Restrictions apply.
www.ietdl.org

ESPRIT [18]. These yield estimates of the Doppler We have presented in this section two possible approaches
frequencies without having to search for the maximum of a to integrate range-distributed target detectors into an
function. operational system whether prior information is available or
not. Further research will be performed int the future on
However, we would like to point out that common spectral comparing and evaluating each method on real data.
estimators will also fulfil this task. Indeed, the periodogram is
sufficient to give performances identical to those of
superresolution techniques. This aspect is shown in Fig. 3a
where the detection probabilities are plotted estimating
3 Ship ISAR imaging and SVM-
Doppler frequencies with ESPRIT (solid line), based automatic ship recognition
rootMUSIC (dash-dotted line) or periodogram (dashed Once the targets have been detected, the objective is to
line) and Fig. 3b the measured false alarm probabilities are compute a high-resolution image of this target in order to
plotted for these three cases. In the case of the extract some features for automatic target recognition.
periodogram, the Doppler frequency estimates are given by
extracting the pr spectrum maxima.

In order to take into account the frequency estimation


3.1 ISAR imaging
error, we define the detection probability as the probability We already proposed an ISAR imaging strategy for ship
that the GLRT is higher than the threshold defined earlier targets in [23] based on selection of imaging time
and that the frequency estimation error is lower than the performed by analysing a series of low-resolution range-
arbitrarily fixed value 1/N, corresponding to the Doppler images. Here we will extend this process by
periodogram frequency resolution. In order to save developing a focusing technique based on tracking the deck
simulation time, the considered target is entirely contained line. The geometry of the acquisition scenario was detailed
in one range cell and is composed of three scatterers of in [23] and corresponds to the observation of a ship
respective normalised Doppler frequencies 0.1, 0.2 and 0.3. animated by roll/pitch and yaw motion observed from a
As we can see, the detection probabilities are almost low flying carrier at long distance. The positions of target
identical whatever spectral estimator is used. However, as scatterers are referred in a (X a , Y a ) plane orthogonal to the
seen in Fig. 3b, false alarm probability depends on the radar line of sight (LOS). We construct the apparent yaw
signal subspace estimate. Indeed, frequency estimation rotation vector VRPa (respectively the apparent roll/pitch
errors or errors on the number of signal components result rotation vector VYa ) corresponding to the projection of the
from the inclusion of clutter components in the steering yaw rotation vector (respectively the roll/pitch rotation
matrix. A more comprehensive study on the influence of vector) on a plane orthogonal to the LOS.
mismatched steering vectors is tackled in [19, 20] and more
recently by De Maio [21]. CFAR behaviour of adaptive We can distinguish the effects of the apparent rotations of
detectors is also studied by De Maio et al. [22] from an the target. Target scatterers at different positions on the
experimental point of view based on real data from the (X a , Y a ) plane have spread contributions to range-Doppler
MIT Lincoln Laboratory Phase-One radar and the images due to the apparent yaw motion. The corresponding
McMaster IPIX radar. range-Doppler image can be related as a top-view-like image

Figure 3 Detection probabilities and the corresponding false alarm probabilities


a Comparison of the detection probabilities on a target composed of three frequencies in one range cell in the case where the frequencies
are known (solid curve with diamonds), ESPRIT-estimated (solid curve), root-MUSIC-estimated (dash-dotted curve) and estimated with the
periodogram (dashed curve)
Pfa ¼ 1022, N ¼ 32, n ¼ 0.5, L ¼ 1
b Corresponding logarithm of computed Pfa

196 IET Signal Process., 2008, Vol. 2, No. 3, pp. 192– 203
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-spr:20070082

Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 22,2010 at 02:37:52 EDT from IEEE Xplore. Restrictions apply.
www.ietdl.org

of the target. Target scatterers at distinct heights have spread R0 V2


contributions to range-Doppler images due to the apparent c¼ (1 þ tan2 F0 )
l
roll/pitch motion. The corresponding range-Doppler image
2R V (14)
can be considered as a side-view-like image of the target [1]. d ¼  0 tan F0
l
This is consistent with the case of spot synthetic aperture e¼f ¼0
radar (Spot-SAR) imaging where the target is stationary
(thus VRPa ¼ 0), whereas the carrier is moving (thus According to (12), the scatterer Doppler frequency is linearly
VYa = 0 due to aspect angle variations) producing a top- varying with time with a slope depending on its distance from
view of the target. In the general case, neither VRPa nor the radar. It represents the ISAR received signal model for
VYa is equal to zero and ISAR processing induces a parts of the target close to sea level, such as the lower bridge
mixture of target side-view and top-view. of the ship target. In fact, as these parts are not high above the
target rotation centre, they are little affected by roll and pitch
In what follows we will only focus on the target side-view motions and can be used as focus points for further
calculation. Thus, the effects of apparent yaw motion are not processing. This yaw motion estimation and compensation
beneficial to the imaging process, but it is judicious to propose act as an autofocus. While standard techniques (Chap. 6 in
a model for its estimation and compensation. We limit the [24]) require a critical accurate range resolution cell selection,
model to short-time observation, enabling us to assume a we here use the entire deck line as a phase reference in a series
constant angular yaw speed V (V is assumed to be the norm of low-resolution range-Doppler snapshots [23].
of VYa ). We do not have to make any assumptions about the
apparent roll/pitch motion. At the beginning of the signal
We apply the yaw motion estimation and correction on a
recording, the target is observed with a look angle F0 . During
radar record of a liquefied gas carrier (Fig. 4). Fig. 5
the recording the look angle becomes F(t) ¼ F0 þ Vt. We
illustrates the focusing process. By measuring a series of
consider a target only animated by yaw motion and a point
Doppler frequency shifts over the deck line 5a, we can fit
scatterer at a distance r̃ of the yaw rotation centre at a distance
the proposed model in Fig. 5b and thus compensate
R0 from the radar. We thus obtain a parametric model of the
accordingly for the phase variation in the original raw data.
distance and Doppler frequency
The two first estimated coefficients of the quadratic form
matrix in (13) should be equal or close to zero if successive
2 estimation of the deck line position is accurate enough.
fD (t) ¼ r~ V sin(F0 þ Vt)
l (10)
d (t) ¼ R0 þ r~ cos(F0 þ Vt) Effects of those phase corrections are illustrated in Fig. 6,
containing three range-Doppler images computed over a 2 s
If we assume that slant range variations are small compared with period with or without corrections. During this integration
the radar range resolution, we find a short-term Doppler time contributions from the ship’s superstructure are
frequency of a target point at a distance r ¼ R0 þ r cos (Fo þ spreading from one side to another of the deck line
Vt) from the radar corresponding to roll motion of the target. Due to this
focusing the range-Doppler image is less blurred, especially
2(r  R0 )V at its rear part and for the ship superstructures.
fD (r, t) ¼ ( tan F0 þ Vt(1 þ tan2 F0 )) (11)
l

This can be reformulated with a matrix representation as follows


[23]

fD (r, t, q ¼ 1) ¼ (r, t, q)M(r, t, q)T (12)


0 1
e a b
M ¼ @a f cA (13)
b c d

and with

V2
a¼ (1 þ tan2 F0 )
l Figure 4 Liquefied petroleum gas (LPG) carrier Henriette
V Kosan sistership (Mette Kosan, picture reproduced by the
b ¼ tan F0
l courtesy of Jim’s Shipping Website [25])

IET Signal Process., 2008, Vol. 2, No. 3, pp. 192 – 203 197
doi: 10.1049/iet-spr:20070082 & The Institution of Engineering and Technology 2008
Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 22,2010 at 02:37:52 EDT from IEEE Xplore. Restrictions apply.
www.ietdl.org

Figure 5 Estimation of Doppler frequency shift


a Variation of measured Doppler shift
b Fitted Doppler shift according to the proposed model

3.2 SVM-based automatic ship in our study, we found that k ¼ 6 gives good results. In
recognition order to reduce the speckle influence, a 3  3 median filter
has been applied on ISAR images. Next, morphological
The last step of the processing sequence is the classification of operations (erosion and dilatation) are applied to close the
detected targets. This task is achieved by comparing the profile and to remove spurious pixels. The final profile is
ISAR images extracted in the previous step to a database of obtained as in [28] by stacking the different contributions
reference profiles; we are then in the context of supervised in each range cell. This makes the extracted profile
learning. Our strategy is to use real ISAR profiles as independent of yaw motion effect that is deck
reference but a database could be constructed or completed line inclination. Fig. 7 shows segmentation results of
with synthetic profiles to circumvent a lack of real data. ISAR image of the LPG carrier Henriette Kosan which is
shown in Fig. 4.
3.2.1 Segmentation: The first step is the extraction of
the ship from the sea clutter background. Although many 3.2.2 Ship databases description and features
noise removal methods are available, we found that global extraction: Although adequate memory storage allows the
thresholding gives satisfactory results. This strategy has use of a huge database of ISAR ship images, we resort to
been adopted by many authors in the ISAR-based ship features-based classification. Indeed, our goal is to build a
classification literature [26 – 30]. fast classification sytem that requires only a small feature space.

The threshold is set from the first- and second-order To characterise ship targets, we considered different kinds
statistics (m, s 2 ) of the sea clutter estimated on the image of features that may traduce information about the
borders, where the signal is assumed free from the target morphology of the target but also the spatial distribution of
influence. Typically it may be chosen as m þ ks 2 . For the RCS. In our study we have not considered the kinematic
example, Gagnon and Klepko [31] choose k ¼ 2 or k ¼ 4; parameters such as the target velocity or acceleration. We

Figure 6 SAR/ISAR range-Doppler maps before and after yaw motion correction
a Beginning of the record
b Two seconds later
c After yaw correction with integration over the first second

198 IET Signal Process., 2008, Vol. 2, No. 3, pp. 192– 203
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-spr:20070082

Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 22,2010 at 02:37:52 EDT from IEEE Xplore. Restrictions apply.
www.ietdl.org

Figure 7 Segmentation of ISAR ship profile of the LPG carrier Henriette Kosan (shown in Fig. 4)
a ISAR image
b ISAR profile
c Normalised profile

considered, among other parameters, the target RCS and its Table 1 Correct classification rate for individual features with
spatial distribution (e.g. the ratio between the energy of the NN classifier applied on DB1 and DB2
front and the total energy), as [26, 32, 30] we considered the
target length, the number of masts and their locations, Features NN, %
attributes of the superstructure (locations of breaks, DB1
percentage of the target, number) and as [26 – 29, 33], we
considered the ISAR profile itself (or an undersampled RCS 82
version) or the Fourier descriptors [34] of its contour. length 53
height 12
Two databases were at our disposal for our study. The first
(DB1) is composed of 11 ship targets: frigates (50 ISAR front energy/total energy ratio 17
images), destroyers (.200 ISAR images), submarines number of masts 8
(150 ISAR images), mine hunters (200 ISAR images), ferries
(45 ISAR images), small boats (22 targets in ISAR images of location of 3 most reflective scatterers 35
small opportunity), cargo (75 ISAR images), freighters space between masts 45
(70 ISAR images), tankers (150 ISAR images), trawlers
(130 ISAR images), tug boats (.200 ISAR images). The masts relative heights 31
second database (DB2) is composed of six classes of targets: height-normalised ISAR profile 83
LPG carriers (33 ISAR images), liquefied natural gas (LNG)
tankers (22 ISAR images), mine hunters (63 ISAR images), DB2
bulk carriers (11 ISAR images), military patrol boats length 65
(21 ISAR images) and trawlers (88 ISAR images). For both
DB1 and DB2, radar data were obtained for standoff ranges height 38
between 10 and 40 nautical miles and sea states ranging front energy/total energy ratio 41
mainly from 2 to 5 (Douglas scale). Each target has been
observed with a certain look-angle range, however not wide number of masts 20
enough to build up a system that takes into account different location of masts 39
look-angles in the classification function. In the learning stage,
targets have thus been included independently of look-angle. masts height 22
number of superstructure breaks 19
In order to select the most discriminating of the above
location of superstructure breaks 38
features, we evaluate the correct classification rate with a
simple nearest-neighbour (NN) classifier. Because of location of maximum 47
operating constraints, features are not exactly the same
% superstructures 35
between DB1 and DB2. Table 1 shows discriminatory
analysis of features for DB1 and DB2 based on NN correct 10 first Fourier descriptors 86
classification rate using the leave-one-out method
height-normalised ISAR profile 87
(LOO) [35]. These tables give an insight into the most

IET Signal Process., 2008, Vol. 2, No. 3, pp. 192 – 203 199
doi: 10.1049/iet-spr:20070082 & The Institution of Engineering and Technology 2008
Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 22,2010 at 02:37:52 EDT from IEEE Xplore. Restrictions apply.
www.ietdl.org

discriminating features (one must keep in mind that these not linearly separable in the reality. Furthermore, it may be
results depend on the databases used and the way they were more judicious to consider nonlinear discriminate
built). functions. The training vectors xi are then mapped into a
higher dimensional space by a function F and a
At last, the features vector we used for ship classification regularisation term is added so that, according to [35], the
is composed of the target length and height, the SVM require the solution of the following optimisation
height-normalised ISAR profile and the ten Fourier problem
descriptors of its contour. We believe that most
information is carried by these features and that other 1 X
min kwk2 þ C ji
features such as mast location or others are redundant. In w 2
i (16)
order to reduce the ship database and to decorrelate the
T
feature vectors, we resort to a principal component analysis. subject to yi (w F(xi ) þ b)  1  ji
We found that a feature subspace dimension set to 10 gives
the best results. C . 0 is the penalty parameter of the error term.
Furthermore, K (x, x0 ) ¼ F(x)T F(x0 ) is called the kernel
3.2.3 SVM classifier: To our knowledge, SVM classifiers function. We have considered a second-order polynomial
are new in the field of automatic ship recognition. They have kernel
already been applied in other radar target recognition
problems such as aircraft in [36] or ground targets in [37, K (x, x0 ) ¼ (xT x0 þ 1)2 (17)
38]. In order to understand the philosophy of the SVM
classifier, we first consider the classification problem Once the discriminate functions have been determined from
between two classes. Let xi [ Rn , i ¼ 1, . . . , K be the the learning dataset, the feature space is then partitioned and
features vectors (with the above considerations, n ¼ 10) we are able to classify new feature vectors.
from the training set and yi [ { 1, 1}, i ¼ 1, . . . , K their
associated class labels. The basic idea is to construct the To extend the SVM classification to a multiclass problem,
optimal separation hyperplane (w, b) so that the one-against-all method is adopted. If M classes are
considered, the principle consists of training M SVM
yi (wT xi þ b)  1 binary classifiers to form a discriminate function
gi , i [ [1, . . . , M], and to decide the class membership of
In the case of linearly separable feature vectors, the previous the feature vector, i.e. whether or not the class i is one of
equation has a solution. However, feature vectors are often the set of remaining classes { j [ [1, . . . , M], j = i}. The

Table 2 Confusion matrix of SVM classification results on DB1

Classifier decision
Fregate Destroyer Submarine Mine Ferry Small Cargo Freighter Tanker Trawler Tug
hunter boats boat
test fregate 87 8 0 4 0 0 0 0 0 0 0
database
destroyer 0 100 0 0 0 0 0 0 0 0 0
submarine 1 0 78 7 0 2 0 0 0 12 0
mine 1 0 2 97 0 0 0 0 0 0 0
hunter
ferry 0 0 0 0 95 0 0 0 5 0 0
small 0 0 22 0 0 56 0 0 0 22 0
boats
cargo 0 5 0 0 0 0 87 8 0 0 0
freighter 2 0 0 0 0 0 12 78 8 0 0
tanker 0 0 1 0 1 0 2 5 91 0 0
trawler 0 0 16 0 0 6 0 0 0 78 0
tug boat 0 0 0 0 0 0 0 0 0 0 100
correct classification rate ¼ 85.0% on DB1 false alarm ¼ 15.4%

200 IET Signal Process., 2008, Vol. 2, No. 3, pp. 192– 203
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-spr:20070082

Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 22,2010 at 02:37:52 EDT from IEEE Xplore. Restrictions apply.
www.ietdl.org

Table 3 Confusion matrix of SVM classification results on DB2

Classifier decision
Patrol boat Trawlers LPG carrier LNG tanker Mine hunter Bulk carrier
test database patrol boat 78 22 0 0 0 0
trawlers 9 91 0 0 0 0
LPG carrier 0 0 100 0 0 0
LNG tanker 0 0 0 100 0 0
mine hunter 0 0 0 0 100 0
bulk carrier 0 0 0 0 0 100
correct classification rate ¼ 94.5% false alarm ¼ 6.9%

decision is made with the following criterion described. We first proposed to use recent HRR-
distributed target detectors that will allow detection of ship
class of x ¼ arg max gi (x) (18) targets in difficult conditions (high sea state, small boat
i[[1,...,M] etc.). Those detectors gives best performance with respect
to low or medium range common target detectors. We
3.2.4 Results on a dataset of real airborne ISAR next proposed an extension of these detectors to estimate
ships images: Tables 2 and 3 show the confusion matrix the target length and Doppler. Two strategies were then
of the classification results on DB1 and DB2, respectively. discussed depending on whether prior knowledge on target
The results reported here have been obtained using the parameters is available or not. Further research will test and
LOO performance estimation technique, which provides an validate these detectors on real data and compare these two
almost unbiased estimate of the classification accuracy [35]. discussed strategies.
SVM classifier is used with a second-order polynomial kernel.
ISAR imaging was presented based on classical and robust
As we witness on DB1 results, the small boats’ class is range-Doppler techniques’ functions. Its aim is here to
confused with the submarine. This may be explained by extract a ship profile based upon a side view of the target
noting that the detectable emerged part of the submarine is and containing the most classifying information. High
approximately the same size of small boats. This class is also resolution is here achieved both by applying a focusing on
confused with trawlers, which is understandable because the the full extent of target deck lines during the observation
shape and lengths of these boats may be very close. Other period and by selecting an imaging moment corresponding
classes are well recognised. The overall correct classification to the greatest Doppler spread.
rate is 85%. Classification results on DB2 are even better,
with an overall correct classification rate of 94.5%. Finally, we developed a classification process. Features
vectors are extracted from ISAR images and SVM classifier
The same classification problems have also been tested is used. We found that SVM gives the best results with
with other classifiers. We compared the Baysesian classifer, respect to other common classifiers and it offers an excellent
the 3NN classifier, and for DB2 the perceptron with three, generalisation capability. Further work should include
five or seven layers and trained with the backpropagation building an SVM-based classification process that takes into
method [39]. For DB1, correct classification rates are account different target look-angles, for example, by simply
67.3% for the Bayesian classifier, 67.6% for the 3NN. In training one classifier per look-angle (or small range of look-
the case of DB2, the results are 85.2, 87.2 and 43.5, 70.2 angles). This should improve classification results.
and 86.5% for the perceptron with three, five or seven
layers, respectively. SVM enables obviously higher
classification rates which may be explained by its nice
generalisation capabilities. In addition, database training 5 References
with SVM is of low computational cost with respect to
methods such as backpropagation. [1] LACOMME P., HARDANGE J.-P., MARCHAIS J.-C., NORMANT E.: ‘Air
and spaceborne radar systems’. (ser. Scitech publishing
Inc., William Andrew Publishing, New York, NY, USA, 2001)
4 Conclusion
Recent results about ship target detection, imaging and [2] GERLACH K.: ‘Spatially distributed target detection in
classification have been described in this paper. A complete non-Gaussian clutter’, IEEE Trans. Aerosp. Electron. Syst.,
system for airborne maritime surveillance has been 1999, 35, (3), pp. 926– 934

IET Signal Process., 2008, Vol. 2, No. 3, pp. 192 – 203 201
doi: 10.1049/iet-spr:20070082 & The Institution of Engineering and Technology 2008
Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 22,2010 at 02:37:52 EDT from IEEE Xplore. Restrictions apply.
www.ietdl.org

[3] CONTE E., MAIO DE, RICCI G.: ‘CFAR detection of distributed [18] PAULRAJ R., KAILATH T. : ‘ESPRIT – a subspace rotation
targets in non-Gaussian disturbance’, IEEE Trans. Aerosp. approach to esimation of parameters of cisoids in noise’,
Electron. Syst., 2002, 38, (2), pp. 612– 621 IEEE Trans. Acoust., Speech, Signal Process., 1986, 34,
pp. 1340 – 1342
[4] BON N., KHENCHAF A., QUELLEC J., GARELLO R.: ‘GLRT-detection
for range- and Doppler-distributed targets in non-Gaussian [19] BLUM R., MCDONALD K.: ‘Analysis of STAP algorithms for
clutter’. In Int. Conf. Radar CIE’06, Shangai, China, 16 – 19 cases with mismatched steering and clutter statistics’,
October 2006 IEEE Trans. Signal Process., Feb. 2000, 48, pp. 301 – 310

[5] BON N., KHENCHAF A., GARELLO R.: ‘GLRT-detection for range- [20] ROBEY F. , FUHRMANN D., KELLY E., NITZBERG R.: ‘A CFAR
and Doppler-distributed targets in non-Gaussian clutter’, IEEE adaptive matched filter detector’, IEEE Aerosp. Electron.
Trans. Aerosp. Electron. Syst., 2008, 44, (2), pp. 678–696 Syst. Mag., 1992, 28, (1), pp. 208 – 216

[6] CONTE E., LONGO M.: ‘Characterisation of radar clutter as a [21] DE MAIO A.: ‘Robust adaptive radar detection in the
spherically invariant random process’, IEE. Proc. F, 1987, presence of steering vector mismatches’, IEEE Trans.
134, (2), pp. 191– 197 Aerosp. Electron. Syst., 2005, 41, (4), pp. 1322 – 1337

[7] WARD K.: ‘Compound representation of high resolution [22] DE MAIO A., FOGLIA G., CONTE E., FARINA A.: ‘CFAR behaviour
sea clutter’, Electron. Lett., 1981, 17, (16), pp. 561 – 563 of adaptive detectors: an experimental analysis’, IEEE
Trans. Aerosp. Electron. Syst., 2005, 41, (1), pp. 233 – 251
[8] CONTE E., LOPS M., RICCI G.: ‘Asymptotically optimum radar
detection in compound Gaussian clutter’, IEEE Trans. [23] HAJDUCH G., CAILLEC L.E., GARELLO R.: ‘Airborne high-
Aerosp. Electron. Syst., 1995, 31, (2), pp. 617– 625 resolution ISAR imaging of ship targets at sea’, IEEE Trans.
Aerosp. Electron. Syst., 2004, 40, (1), pp. 378– 384
[9] GINI F., FARINA A.: ‘Vector subspace detection in compound-
Gaussian clutter part I: survey and new results’, IEEE Aerosp. [24] CARRARA W.G., GOODMAN R.S., MAJEWSKI R.M. : ‘Spotlight
Electron. Syst. Mag., 2002, 38, (4), pp. 1295–1311 synthetic aperture radar: signal processing algorithms’,
ser. ‘remote sensing library’ (Artech House, Boston, MA,
[10] GINI F., FARINA A.: ‘Vector subspace detection in compound- 1995)
Gaussian clutter part II: performance analysis’, IEEE Aerosp.
Electron. Syst. Mag., 2002, 38, (4), pp. 1312–1323 [25] Jim’s Shipping Website: http://www.jimsshippingwebsite.
co.uk/bristol.htm
[11] SWERLING P.: ‘Probability of detection for fluctuating
targets’, IEEE Trans. Inf. Theory, 1960, 6, (2), pp. 269– 308 [26] MUSMAN S., KERR D., BACHMANN C.: ‘Automatic recognition
of ISAR ship images’, IEEE Trans. Antennas Propag., 1996,
[12] NAKAGAMI M.: ‘The m-distribution, a general formula of 32, (4), pp. 1392– 1404
intensity distriubtion of rapid fading’. (Pergamon,
W.G. Hoffman, Oxford, UK, 1960) [27] RICE F. , COOKE T., GIBBINS D. : ‘Model based ISAR ship
classification’, Digit. Signal Process., 2006, 16, (5),
[13] BLACKMAN S.S.: ‘Multiple-target tracking with radar pp. 628– 637
application’ (Artech House, 1986)
[28] MENON M.: ‘An automatic ship classification system for
[14] WATTS S.: ‘Cell-averaging CFAR gain in spatially ISAR imagery’. Proc. Applications and Science of Artificial
correlated K-distributed sea clutter’, IEE Proc., Radar Neural Networks, Proc. SPIE, 1995, 2492, pp. 373– 388
Sonar Navig., 1996, 143, (5), pp. 321 – 327
[29] MAKI A., FUKUI K.: ‘Ship identification in sequential ISAR
[15] ANDERSH D., HAZLETT M., LEE S., REEVES D., SULLIVAN D., CHU Y.: imagery’, Mach. Vis. Appl., 2004, 15, (3), pp. 149– 155
‘Xpatch: a high-frequency electromagnetic scattering
prediction code and environment for complex three- [30] YUAN C. , CASASENT D. : ‘Composite filters for inverse
dimensional objects’, IEEE Antennas Propag. Mag., 1994, synthetic aperture radar classification’, Opt. Eng., 2002,
36, (1), pp. 65– 69 41, (1), pp. 94– 104

[16] RISSANEN J. : ‘Modeling by shortest data description [31] GAGNON L., KLEPKO R.: ‘Hierarchical classifier design for
length’, Automatica, 1978, 14, pp. 465– 471 airborne SAR images of ships’. SPIE Proc. Conf. Automatic
Target Recognition VIII, Orlando, 1998
[17] SCHMIDT R.: ‘Multiple emitter location and signal
paramter estimation’, IEEE Trans. Antennas Propag., 1986, [32] MORUZZIS M., SAULAIS P., TAT T. , HUEI T.: ‘Automatic
34, pp. 276 – 280 target classification for naval radar’. Int. Conf.

202 IET Signal Process., 2008, Vol. 2, No. 3, pp. 192– 203
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-spr:20070082

Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 22,2010 at 02:37:52 EDT from IEEE Xplore. Restrictions apply.
www.ietdl.org

Radar Systems, RADAR’04, Toulouse, France, 12 – 22 [36] RADOI E., TOTIR F., QUINQUIS A., ANTON L.: ‘Superresolution
October 2004 imagery based SVM classification of radar targets’. EUSAR, 2006

[33] KNAPSKOG A.: ‘Automatic classification of small ships in [37] BRYANT M., GARBER F. : ‘SVM classifier applied to the
ISAR images using 3D models and silhouette matching’. MSTAR public data set’. Algorithms for Synthetic Aperture
EUSAR, 2006 Radar Imagery VI - Proc. SPIE, 1999, 3721, pp. 355 – 360

[34] ZAHN C., ROSKIES R.: ‘Fourier descriptors for plane [38] ZHAO Q., PRINCIPE J.: ‘Support vector machines for SAR
closed curves’, IEEE Trans. Comput., 1972, 21, (3), automatic target recognition’, IEEE Trans. Aerosp.
pp. 269– 281 Electron. Syst., 2001, 37, (2), pp. 643– 654

[35] DUDA R., HART P., STORK D.: ‘Pattern classification’ (John [39] THEODORIDIS S., KOUTROUMBAS K.: ‘Pattern recognition’
Wiley & Sons, 2000) (Elsivier Academic Press, 2003, 2nd edn.)

IET Signal Process., 2008, Vol. 2, No. 3, pp. 192 – 203 203
doi: 10.1049/iet-spr:20070082 & The Institution of Engineering and Technology 2008
Authorized licensed use limited to: FUDAN UNIVERSITY. Downloaded on March 22,2010 at 02:37:52 EDT from IEEE Xplore. Restrictions apply.

Potrebbero piacerti anche