Professional Documents
Culture Documents
Application of The Radon Transform To Detect Small-Targets in Sea Clutter
Application of The Radon Transform To Detect Small-Targets in Sea Clutter
Application of The Radon Transform To Detect Small-Targets in Sea Clutter
org
ISSN 1751-8784
Abstract: The authors present a novel and heuristic approach for the detection of low radar cross-section targets
in high-resolution sea clutter. The proposed technique is based on the application of the Radon transform to
range–time matrices formed by column-wise storage of consecutive range profiles. The objective of this paper
is 2-fold: to analyse the effect of the transform on real high-resolution sea clutter and to describe a detection
scheme based on the insight obtained. The proposed technique emulates the behaviour of traditional motion
target detection algorithms without the need for reliable Doppler information. It also constitutes a powerful
non-coherent integration strategy of the target’s energy along its specific path on the range – time plot. The
performance of the detection technique has been tested against real high-resolution sea clutter data, acquired
at the south coast of Spain with an in-house developed continuous wave linear frequency modulated
millimetre-wave radar system. Monte Carlo simulations show a significant improvement over the conventional
cell averaging constant false alarm rate schemes.
IET Radar Sonar Navig., 2009, Vol. 3, Iss. 2, pp. 155 – 166 155
doi: 10.1049/iet-rsn:20080123 & The Institution of Engineering and Technology 2009
www.ietdl.org
experimental sea clutter amplitude probability density Statistical analysis was carried out with the data used for
functions (PDFs) [5 – 9]. this research. A good agreement was found with the
compound-Gaussian model following the statistical analysis
Conventional radar detection chains were usually designed techniques used in earlier publications [5, 7]. Specifically,
to detect targets embedded in Gaussian-distributed and both the experimental amplitude PDF and the measured
uncorrelated noise backgrounds. Consequently, there is correlation properties are compatible with the compound
unacceptable performance degradation when used against model. Among the most commonly used bi-parametric
high-resolution sea clutter [13]. For example, it is known PDFs, the best fit is obtained by the generalised
that the behaviour of conventional cell averaging constant K-distribution with log-normal texture, and the existence
false alarm rate (CA-CFAR) is degraded by noise processes of the two components, the so-called speckle and texture,
with such long-tailed amplitude PDFs, and both temporal has also been verified. The complete statistical analysis of
and spatial non-negligible correlations. the available data sets is the topic of a future publication.
Recently, much effort has been devoted to the definition of The remainder of the paper is organised as follows. Section
detection schemes robust against compound-Gaussian noise 2 briefly reviews the Radon transform, some of its properties
[14 – 16]. Although the reported performance of some of and implementation issues. Section 3 begins with a description
them is good [17], the detection of small targets in sea of the proposed detection technique, highlighting the effect
clutter is not yet a closed problem. In this paper, a Radon of the Radon transform on real sea clutter data. Afterwards,
transform based approach for the detection of small targets two different automatic detection schemes are detailed.
in sea clutter is proposed and its performance is tested Section 3 is an extension of [19]. An evaluation of the
against real sea clutter data. performance of the proposed detection schemes is included
in Section 4. Finally, conclusions are drawn in Section 5.
Experimental low grazing angle sea clutter time series were
recorded at the south coast of Spain with an in-house
developed high-resolution continuous wave linear frequency 2 Radon transform
modulated (CWLFM) radar demonstrator [18]. The main
The Radon transform of a continuous two-dimensional
characteristics of the radar system are detailed in Table 1.
function g(x,y) can be defined as [20]
The radar’s maximum transmitted bandwidth was 2 GHz
and thus a range resolution better than 10 cm was achieved. ðð
The data set comprises measurements at different R(r, u) ¼ g(x, y)d(r x cos u y sin u)
configurations, with varying azimuth angles, transmitted D
Table 1 Radar characteristics The most outstanding and well-known application of the
Radon transform is in the field of medical imaging,
Radar characteristics providing the theoretical principles of computerised
tomography. However, it has been extensively used for
Type Continuous wave linear
pattern recognition, as it is a particularly well-suited
frequency modulated
technique for line-detection tasks. It can be easily shown
Tx. frequency 28– 30 GHz that for each line feature in the image domain, there will be
an associated peak in the Radon space. This fact is
Bandwidth 2 GHz (Max.)
exploited by Radon transform based line-detection
Waveform repetition 3000 Hz (Max.) algorithms.
frequency
Only certain functions can be analytically Radon
Tx. power 30 dBm
transformed and therefore several discrete approximations
156 IET Radar Sonar Navig., 2009, Vol. 3, Iss. 2, pp. 155– 166
& The Institution of Engineering and Technology 2009 doi: 10.1049/iet-rsn:20080123
www.ietdl.org
IET Radar Sonar Navig., 2009, Vol. 3, Iss. 2, pp. 155 – 166 157
doi: 10.1049/iet-rsn:20080123 & The Institution of Engineering and Technology 2009
www.ietdl.org
intrinsic wide spectrum of sea clutter severely degrade the where K1 is a parameter of the detection chain and mean
performance of Doppler-based MTD techniques. (R(r, u)) and std (R(r, u)) are the global mean and standard
deviation of the Radon transformed range–time matrix. The
Detailed visual inspection of Fig. 1a shows that wave- chosen expression for T1 is a simple outlier detection
crests have low radial speeds (slope of the line-like features technique [32]. The value for K1 is experimentally chosen:
on the range – time plot). Targets having different radial the peaks associated with wave-crests should be above T1 , as
speeds can be easily simulated and inserted into real sea they are used in the following steps of the detection chain.
clutter data backgrounds. Certain speed ranges will give
range– time plots with different slopes than wave-crests. After the application of the first threshold, each wave-crest
The detection scheme based on the Radon transform will (or target) has an associated cluster of points. Each cluster of
exploit this difference to avoid false alarms from target-like first threshold crossings is a unique ‘event’. There are many
wave-crest radar returns. different clustering techniques in the image processing
literature. The proposed detection chain looks for
The preceding discussion constitutes a heuristic ‘8-connected’ points in the thresholded Radon image using
motivation for the use of the Radon transform for the a connected component labelling algorithm [33]. After the
detection of targets in sea clutter. The idea is further clustering step, it is possible to calculate statistical properties
supported by previous and successful applications of this of the points of each event. The calculated statistical
transform in the fields of radar signal processing, computer properties should make possible the discrimination between
vision and pattern recognition. The Radon transform based the two kinds of events: wave-crests and targets. The
track-before-detect system described in [27] or the ship proposed detector labels each event with two values:
wake detection systems proposed in [28 – 30] are good
examples that clearly demonstrate the potential of this 1. The maximum value of the Radon transform of the points
approach for the detection of linear structures in noise. of each event.
Moreover, it can be shown from a detection theory point of
view that the optimal detector of straight patterns in noise- 2. The u coordinate at which the maximum occurs.
corrupted images is based on the Radon transform [31].
Therefore there is a statistical justification for the use of the The output of the clustering block is a list of events for the
transform in line-like feature-detection problems that goes current range – time scene. The final detection is performed
beyond the intuition prompted by the analysis of the in an amplitude against u coordinate space. The decision
available experimental sea clutter data sets. boundary can be set taking into account the following
considerations:
3.1 Radon transform-based basic
1. Targets with radial speeds (u coordinates) different from the
detection scheme average of wave-crests should be detected regardless of their
Fig. 2 shows the block diagram of a basic version of the ratio target amplitude to wave-crest peak amplitude. This
proposed detection system. Consecutive range profiles are should allow for the detection of small targets below sea clutter.
stored column-wise to form a range–time matrix. A direct
discrete Radon transform with nearest-neighbour 2. Targets with u coordinates similar to those of wave-crest
interpolation is applied to the range– time image. A first events will be detected just if their ‘amplitude to wave-crest
threshold is then used to detect peaks on the Radon peak amplitude ratio’ is high enough.
transformed image. This first threshold is calculated as
3.2 Effect on experimental sea clutter data
T1 ¼ meanðRðr, uÞÞ þ K1 stdðRðr, uÞÞ
The strategy explained in the previous section can be
simulated by inserting synthetic low-amplitude targets,
and applied with different speeds, into the available real sea clutter data
sets. Fig. 3a shows a representative example of a scene to
be processed. 256 range profiles are stored as a range–time
R ri , uj , T1 ) R ri , uj ¼ 0
matrix. There are 256 range bins in each range profile with
158 IET Radar Sonar Navig., 2009, Vol. 3, Iss. 2, pp. 155– 166
& The Institution of Engineering and Technology 2009 doi: 10.1049/iet-rsn:20080123
www.ietdl.org
a range resolution of 0.16 m. The fm of the radar is 1 kHz and Radon transform. The limited dynamics of targets can be
therefore the integration time is 256 ms. The following exploited to bound the required u interval: most angles
equation can be used to calculate u as a function of the have no physical meaning for maritime surveillance radars.
radial speed of the synthetic target: This fact reduces the computational requirements of the
proposed techniques making real-time implementations
1 Vr 0:5144 possible. A carefully chosen step for u can also minimise
u ¼ 90 + tan the computational load of the Radon transform block.
fm Dr
where Vr is the radial speed (knots), fm is the frequency (Hz) In the scene shown in Fig. 3a, two synthetic targets have
of the modulating waveform of the CWLFM radar system been inserted. Their simulated speeds are þ15 and þ50
and Dr is the range resolution (m). knots. It is difficult to detect them by visual inspection
because their amplitude is below sea clutter wave-crests.
The maximum expected speed of targets in a specific Both targets are easily identified in the Radon transformed
maritime surveillance environment can be set. 50 knots is a image (Fig. 3b). The output of the basic detection scheme
reasonable value, which corresponds to an angle of +98 is shown in Fig. 4a, in which events extracted from the
around 908 on the range – time plot. A horizontal line in scene are plotted in an amplitude against angle graph. Both
the range – time plot is a static target (u ¼ 908). It is not targets could be easily detected if an appropriate detection
necessary to calculate all of the projection angles of the mask on the amplitude against u coordinate space is used.
IET Radar Sonar Navig., 2009, Vol. 3, Iss. 2, pp. 155 – 166 159
doi: 10.1049/iet-rsn:20080123 & The Institution of Engineering and Technology 2009
www.ietdl.org
160 IET Radar Sonar Navig., 2009, Vol. 3, Iss. 2, pp. 155– 166
& The Institution of Engineering and Technology 2009 doi: 10.1049/iet-rsn:20080123
www.ietdl.org
IET Radar Sonar Navig., 2009, Vol. 3, Iss. 2, pp. 155 – 166 161
doi: 10.1049/iet-rsn:20080123 & The Institution of Engineering and Technology 2009
www.ietdl.org
clear contrast with the earlier works, which relied on this small target detection system that includes a profile
parameter for artefact rejection [29]. The parameter that analysis processing block.
provides the best class discrimination potential is the slope
of the peak, which can be approximated by the difference 3.4 Radon transform based detection
between consecutive profile samples. Fig. 7b demonstrates
that if the chosen feature is the maximum of the absolute
system using profile analysis
value of the differentiated normalised profile, the Fig. 8 shows the proposed Radon transform based detection
discrimination between the two classes is possible. The system. Two processing blocks have been added to reduce the
next section describes a modified Radon transform based effect of artefacts on the performance of the technique. The
162 IET Radar Sonar Navig., 2009, Vol. 3, Iss. 2, pp. 155– 166
& The Institution of Engineering and Technology 2009 doi: 10.1049/iet-rsn:20080123
www.ietdl.org
Figure 8 Block diagram of the Radon transform based detection system using profile analysis
first block performs morphological operations in the Radon simulations have been performed for a fixed probability of
space after the application of the first threshold. false alarm of 0.001 (Fig. 9). Fig. 9a shows the probability of
Morphologic erosion is applied followed by a dilation using detection as a function of signal-to-clutter ratio. One
an appropriate structuring element [32]. This operation is a thousand realisations of random speed targets have been
morphological opening and eliminates structures from the used to estimate each point of the curve. The simulation
thresholded Radon space with dimensions smaller than the assumes constant velocity targets. In a real maritime
structuring element. The size and shape of the element has surveillance scenario, the Radon transform will provide a
been chosen to mimic the Radon space footprint of a low- first-order approximation to the target trajectory in the
amplitude target. Clusters of first threshold crossings range– time representation. The location of peaks in the
smaller than the defined structuring element are artefacts Radon space is going to be related to the first-order
and therefore removed. Experimental tests have determined approximated and time-averaged dynamical behaviour of the
that a 3 pixel wide line in the u direction is an appropriate target. This is the approach of previous works such as [27].
choice for the structuring element of the morphological The consequence of the constant velocity assumption during
processing block. This block prevents the clustering the dwell time is a slight uncertainty in the estimated
algorithm from generating events for certain artefacts and dynamics that can be neglected for detection purposes.
reduces the computational load of the profile analysis block.
For comparison, Fig. 9a also shows the performance of
The profile analysis block processes the event list produced conventional CA-CFAR processors:
by the clustering algorithm. For each event, a profile in the r
coordinate is extracted across its maximum value. The profile 1. Square law detector followed by a CA-CFAR. Two guard
is limited to a symmetrical neighbourhood around the peak cells and a variable number of reference cells (Nr).
value, normalised, and the feature described in the previous
section is calculated. Each event is labelled with a feature 2. Square law detector followed by non-coherent integration
vector defined as of the whole scene and a CA-CFAR. Variable number of
guard and reference cells (Nr).
xi ¼ Ai , ui , fi
For all the simulated processing chains, a conventional
where Ai is the maximum amplitude of the event, ui is the methodology in radar performance analysis has been used:
angular coordinate of the peak and fi is the maximum of
the absolute value of the normalised profile differentiated.
The detection step can take place in a feature space defined 1. Set the scale factor (CA-CFARs) or detection mask
by vectors xi . A visual inspection of Fig. 7b suggests a more (Radon-based detectors) by Monte Carlo simulation of
convenient approach based on the application of a entire data sets corresponding to consecutive data
threshold K2 to fi . Most artefacts can be rejected by setting acquisitions. Because of the amount of data available, a
a value between 0.4 and 0.6 for this threshold. The analysis fairly high probability of false alarm has been used (0.001).
of several different experimental sea clutter data sets has
demonstrated that a value of K2 ¼ 0.55 is a good choice. 2. With the experimentally chosen scale factor or detection
mask, perform the Monte Carlo simulation of the
With the two new processing blocks, it is possible to lower probability of detection for each signal-to-clutter ratio.
the detection mask for the same false alarm rate when
compared with the basic detection scheme, thus improving The main result is an improvement of almost 8 dB when
the performance of the detection technique. Monte Carlo compared with the conventional CA-CFAR schemes. The
simulations of the proposed detection system are included previously discussed effect of the chosen value for the first
in the following section. threshold (K1) can also be appreciated in Fig. 9a. Low
values for this parameter improve the detection capabilities
of the system for low SCR targets at a cost of a slight
4 Results performance degradation for average SCR targets. Fig. 9a
The availability of experimental sea clutter data sets has also includes a comparison with non-coherent integration
allowed a realistic performance evaluation of the Radon followed by a CA-CFAR. Two curves are shown for this
transform based detection techniques. Monte Carlo case, the INT-CA-CFAR performance against a static
IET Radar Sonar Navig., 2009, Vol. 3, Iss. 2, pp. 155 – 166 163
doi: 10.1049/iet-rsn:20080123 & The Institution of Engineering and Technology 2009
www.ietdl.org
target (the best case for this detector) and against a þ5 knot The main requirement for the proposed detection system is
target. The strong performance degradation of the INT-CA- to make the detection of small targets possible, provided that
CFAR against moving targets, a consequence of range cell their radial speed is different enough from wave-crests.
migration, is clearly demonstrated by the þ5 knot curve. Fig. 9b shows that targets with slow radial speeds (but of
The proposed Radon transform based scheme can be opposite sign when compared with wave-crests) are
regarded as a range alignment process before a non- detected. The probability of detection for þ1 knot target is
coherent integration and therefore is inherently robust 60%, and reaches its nominal value at þ2 knot.
against range cell migration. This is one of the reasons for
the improvement over the INT-CA-CFAR that is shown Monte Carlo simulations for single range bin targets
in Fig. 9a. clearly show that the proposed approach is suitable for the
164 IET Radar Sonar Navig., 2009, Vol. 3, Iss. 2, pp. 155– 166
& The Institution of Engineering and Technology 2009 doi: 10.1049/iet-rsn:20080123
www.ietdl.org
detection of very small targets in sea clutter. The detection of [3] HAYKIN S., BAKKER R., CURRIE B.W.: ‘Uncovering nonlinear
periscopes is one of the tasks for which the described dynamics – the case study of sea clutter’, Proc. IEEE, 2002,
processing scheme is particularly well suited and for which 90, (5), pp. 860– 881
the assumptions made for the simulations are reasonable.
Nevertheless, it is important to stress that Radon transform [4] UNSWORTH C.P., COWPER M.R., MCLAUGHLIN S., MULGREW B.: ‘Re-
based systems can cope with the detection of range-extensive examining the nature of radar sea clutter’, IEE Proc. Radar
targets. Range-extensive targets will have a ‘flat peak’ Signal Process., 2002, 149, (3), pp. 105 – 114
signature in the Radon space [34]. As explained in the
previous sections, the r profile analysis is limited to the [5] FARINA A., GINI F., GRECO M.V., VERRAZZANI L.: ‘High resolution
neighbourhood of each peak. To account for ‘flat peak’ sea clutter data: statistical analysis of recorded live data’,
target signatures, only minor modifications are required: the IEE Proc. Radar Signal Process., 1997, 144, (3), pp. 121– 130
proposed r profile analysis should be performed on a longer
neighbourhood to include the rising and falling edges of the [6] NOHARA T.J., HAYKIN S.: ‘Canadian East Coast radar trials
range-extended target r profile inside the processing window. and the K-distribution’, IEE Proc. Radar Signal Process.,
1991, 138, (2), pp. 80– 88
Finally, it is necessary to state that the technique has good
performance for other operative configurations (range [7] CONTE E., DE MAIO A., GALDI C.: ‘Statistical analysis of real
resolution and dwell time). There is a link between the clutter at different range resolutions’, IEEE Trans. Aerosp.
required range resolution and integration time, as targets or Electron. Syst., 2004, 40, (3), pp. 903– 918
wave-crests should move across range bin boundaries
during the dwell time. As a consequence, and to achieve [8] GRECO M., BORDONI F., GINI F.: ‘X-band sea-clutter
the performance described in this section, it would be nonstationarity: influence of long waves’, IEEE J. Oceanic
possible to divide the required dwell time by 2 if the range Eng., 2004, 29, (2), pp. 269– 283
resolution is doubled. For lower transmitted bandwidths,
the integration time should be longer to keep the range [9] GRECO M., GINI F. , RANGASWAMY M. : ‘Non-stationarity
resolution to dwell time ratio constant. analysis of real X-band clutter data at different
resolutions’. IEEE Conf. Radar, 2006, pp. 7 – 24
5 Conclusions [10] WARD K.D.: ‘Compound representation of high
In this paper, a novel approach for the detection of low RCS resolution sea clutter’, Electron. Lett., 1981, 17, p. 561
targets in maritime surveillance systems is proposed. Two
Radon transform based detection schemes are detailed and [11] DI BISCEGLIE M. , GALDI C. : ‘Random walk based
their performance is tested against real high-resolution sea characterisation of radar backscatter from the sea
clutter. The detection is performed in an amplitude against surface’, IEE Proc. Radar Sonar Navig., 1998, 145, p. 216
angle (speed) space emulating traditional MTD processors.
Monte Carlo simulations show an improvement over the [12] JAKEMAN E., PUSEY P.: ‘A model for non-Rayleigh sea echo’,
conventional CA-CFAR detection schemes. Ongoing IEEE Trans. Antennas Propag., 1976, 24, (6), pp. 806– 814
research is focused on the statistical analysis of Radon
transformed high-resolution sea clutter and on automatic [13] WATTS S., BAKER C.J., WARD K.D.: ‘Maritime surveillance
and adaptive strategies for the definition of the detection radar. II: Detection performance prediction in sea clutter’,
mask. IEE Proc. Radar Signal Process., 1990, 137, (2), pp. 63– 72
IET Radar Sonar Navig., 2009, Vol. 3, Iss. 2, pp. 155 – 166 165
doi: 10.1049/iet-rsn:20080123 & The Institution of Engineering and Technology 2009
www.ietdl.org
IEEE Trans. Aerosp. Electron. Syst., 2005, 41, (1), [26] FREDERICK M.T., VANDERHORN N.A., SOMANI A.K.: ‘Real-time
pp. 233– 251 H/W implementation of the approximate discrete Radon
transform’. 16th IEEE Int. Conf. Application-Specific
[18] ASENSIO-LOPEZ A., BLANCO-DEL-CAMPO A., GISMERO-MENOYO J. , Systems, Architecture Processors, 2005 (ASAP 2005),
ET AL .:
‘High range-resolution radar scheme for imaging 23– 25 July 2005, pp. 399 – 404
with tunable distance limits’, Electron. Lett., 2004, 40,
(17), pp. 1085 – 1086 [27] CARLSON B.D., EVANS E.D., WILSON S.L.: ‘Search radar detection
and track with the Hough transform. I. System concept’, IEEE
[19] CARRETERO-MOYA J. , GISMERO-MENOYO J. , ASENSIO-LOPEZ A. , Trans. Aerosp. Electron. Syst., 1994, 30, (1), pp. 102–108
BLANCO-DEL-CAMPO A.:‘Small-target detection in sea clutter
based on the Radon transform’. Proc. Int. Conf. Radar [28] MAGLI E., OLMO G., LO PRESTI L.: ‘On-board selection of
2008, pp. 610 – 615 relevant images: an application to linear feature recognition’,
IEEE Trans. Image Process., 2001, 10, (4), pp. 543–553
[20] TOFT P.: ‘The radon transform – theory and
implementation’. PhD thesis, Department of Mathematical [29] COPELAND A.C., RAVICHANDRAN G., TRIVEDI M.M.: ‘Localized Radon
Modelling, Technical University of Denmark, 1989 transform-based detection of ship wakes in SAR images’, IEEE
Trans. Geosci. Remote Sens., 1995, 33, (1), pp. 35–45
[21] BEYLKIN G. : ‘Discrete radon transform’, IEEE Trans.
Acoust. Speech Signal Process., [see also IEEE Trans. [30] REY M.T., TUNALEY J.K., FOLINSBEE J.T., JAHANS P.A., DIXON J.A., VANT
Signal Process.] 1987, 35, (2), pp. 162– 172 M.R.: ‘Application of radon transform techniques to wake
detection in seasat-asar images’, IEEE Trans. Geosci.
[22] KELLEY B.T., MADISETTI V.K.: ‘The fast discrete Radon Remote Sens., 1990, 28, (4), pp. 553 – 560
transform. I: Theory’, IEEE Trans. Image Process., 1993, 2,
(3), pp. 382– 400 [31] NERI A.: ‘Optimal detection and estimation of straight
patterns’, IEEE Trans. Image Process., 1996, 5, (5),
[23] GOTZ W.A., DRUCKMULLER H.J. : ‘A fast digital radon pp. 787– 792
transform – an efficient means for evaluating the Hough
transform’, Pattern Recog., 1996, 29, (4), pp. 711– 718 [32] WOODS R.E., GONZALEZ R.C.: ‘Digital image processing’
(Prentice Hall, 2002, 2nd edn.)
[24] YANG D.: ‘Fast discrete radon transform and 2-D discrete
Fourier transform’, Electron. Lett., 1990, 26, (8), pp. 550–551 [33] HARALICK , ROBERT M., SHAPIRO L.G.: ‘Computer and robot
vision, vol. I’ (Addison-Wesley, 1992), pp. 28– 48
[25] BRADY M.L., YONG W.: ‘Fast parallel discrete approximation
algorithms for the radon transform’. Proc. Fourth Annual [34] ZHANG Q., COULOIGNER I.: ‘Accurate centerline detection and
ACM Symp. Parallel Algorithms and Architectures, San line width estimation of thick lines using the radon transform’,
Diego, CA, 1992, pp. 91 – 99 IEEE Trans. Image Process., 2007, 16, (2), pp. 310–316
166 IET Radar Sonar Navig., 2009, Vol. 3, Iss. 2, pp. 155– 166
& The Institution of Engineering and Technology 2009 doi: 10.1049/iet-rsn:20080123