Professional Documents
Culture Documents
By
2003
______________________
______________________
Committee on
Final Examination
___________________________
___________________________
___________________________
___________________________
___________________________
High Range Resolution (HRR) Profiles is presented and also a number of ways are
investigated for target detection using Synthetic Aperture Radar (SAR) images.
for sequential High Range Resolution (HRR) radar signatures. The proposed hybrid
proposed hybrid approach, each HRR test profile is first scored by ETMF that is then
followed by independent HMM scoring. The first ETMF scoring step produces a limited
number of “most likely” models that are target and aspect dependent. These reduced
numbers of models are then used for improved HMM scoring in the second step. Finally,
the individual scores of ETMF and HMM are combined using Maximal Ratio Combining
to render a classification decision. Classification results are presented for the MSTAR
that employs physical and statistical models is developed and presented. This joint
iii
physics/statistics based technique generates images that have many of the “blob-like” and
“spiky” clutter characteristics of UWB radar data in forested regions while avoiding the
UWB radar target detection are investigated using the results of this simulation process.
iv
CONTENTS
1: Introduction 1
2.1.2: Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
v
2.1.3: Alignment of HRR Profiles in Range . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.2.1: Framing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.2.2.2: Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
vi
2.3.2.1: Necessity of developing modified hybridization in ETMF ATR 56
2.4.3.1: Framing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.4.3.2: Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.4.5: Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
vii
3.2.1.1: Description of MSTAR Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.2.3: Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
viii
4.5.4.2.2: Performance Comparison in Self-Training mode . . . . 129
Bibliography 141
ix
List of Figures
2.2 Distribution of Singular values for MSTAR target T72, 1000 sector . . . 23
2.4 Observation and Template Profiles are shown in shaded and blank 27
2.9 Two states Hidden Markov Model with two output symbols, V1 and 32
V2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
of the joint event that the system is in state Si at time t and state Sj at 43
time t+1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
x
2.13 Block Diagram of a Designed HMM recognizer . . . . . . . . . . . . . . . . . 47
2.17 In 100 aspect case, this plot shows the HMM recognition rate with 60
2.18 This figure is used to determine the most effective W1/ W2 so that the 62
performances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
xi
2.26 ROC curves for Probability of Declaration vs Conditional Probability 82
xii
4.6 Sample UWB SAR Simulation test Image . . . . . . . . . . . . . . . . . . . . . . . 119
4.10 UWB SAR simulation image after performing Euclidean masking 125
operation in TF-SVD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.11 Final UWB SAR simulation image after performing TF-SVD . . . . . . . 126
train mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
scale) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
xiii
List of Tables
VII Confusion matrix for HMM with three profile average testing . . . . 70
VIII Confusion matrix for Hybrid algorithm with three profile average 70
testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
xiv
XV Confusion matrix (Unknown rejection threshold about 0.6) for 76
Hybrid classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Hybrid classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
0.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Pd = 0.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Pfa = 0.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
xv
ACKNOWLEDGEMENTS
goal to pursue graduate studies. First I would like to thank my parents, Goutam K. Paul
and Sipra Paul, for their constant support and encouragement. They have made lots of
I would like to thank Professor Arnab K. Shaw, WSU, and Dr. Atindra K. Mitra,
also like to thank Dr. Kefu Xue and Dr. Fred Garber for agreeing to be on my thesis
committee.
I would like to acknowledge my friends, Koel Das and Sivaram Bandaru for
Lastly, I would wish to thank all the faculty members of the Electrical
Engineering Department at Wright State University for their generous help and
xvi
1: INTRODUCTION
The objective of Automatic Target Recognition (ATR) algorithms is to correctly identify
an unknown target from sensed radar signatures [1-4], whereas in target detection case,
the requirement is to detect target from clutter. The need for ATR and target detection
technology is evident from various “friendly fire” incidents. The most popular algorithm
for ATR is the template-matching algorithm. Given a sensed signature from an unknown
target, the ATR systems compare the observed signatures with a set of stored target
hypotheses. The target decision is based on some form of optimum similarity between the
observed signature and one of the stored targets. Template based ATR provides
encouraging results as demonstrated in the work of Novak, et al. [5], Mirkin [6] and
many others [7-9]. Whereas in target detection case, the classifier is trained to determine
the threshold, which is a discriminant factor between target and clutter. Based on this
threshold the classifier will perform target detection while nullifying clutter.
In the next subsections a brief review of ATR/target detection and its background and
The present era of limited warfare demands precision strikes for reduced risk and cost
efficient operation with minimum possible collateral damage. In order to meet such
1
exacting challenges, Automatic Target Recognition (ATR)/Target Detection capability is
becoming increasingly important to the Defense community. The overall goals are to
analyze image data using digital computers in order to detect, classify and recognize
target signatures automatically, i.e., with minimum possible human assistance. The image
data for processing may be generated by one of many possible imaging sensors including
one of the most challenging among current research problems because the system
developers have little control over the possible target scenario and the operational
imaging condition [14-17]. Also, compared to the diversity of possible images during
algorithms may have to deal with intelligent adversary attempting to defeat the system, as
into two general areas: Moving Target Indication (MTI) and Synthetic Aperture Radar
(SAR) [18-22]. The original purpose for developing these radar technologies had been to
achieve all weather and all day/night imaging, i.e., to transcend traditional photographic
camera based imaging that must rely on sunlight and is susceptible to clouds, fog or
precipitation.
Most surface and airborne radar systems operate in an environment where the clutter
return obscures targets of interest [23]. If the target is moving relative to the clutter it is
2
possible to filter out the undesired clutter return by exploiting the differential doppler
frequency shift produced by relative target to clutter radial motion. Systems following
MTI has the capability to detect target reflections [24] having differential radial
motion with respect to the clutter. The clutter causing background may be either terrain,
sea, weather or chaff [25-26]. MTI’s are operated with either fixed based or a moving
i.e. the radar is surface based, flying over terrain through possible weather disturbances.
In such an event, MTI rejects the returns from terrain and weather while retaining the
return from the aircraft. This property gives it good detection capabilities for air borne
targets. In cases where the target is surface based, as in Air to Ground ATR application,
the ground clutter are stronger than the expected target return. The ground clutter
extends out to a range where terrain features that cause the clutter are masked due to
earth's curvature. In such cases, the ground clutter extends to the full operating range of
MTI is a mature radar technology that allows airborne sensors to survey large
areas of land and it has coarse target detection and range determination capabilities. It
makes use of target movement for image formation and hence, it is highly effective for
distinguishing moving targets from ground clutter. However, a major drawback of the
3
1.1.3: Synthetic Aperture Radar (SAR)
Although the major emphasis of this thesis is to utilize HRR profiles, as described in the
previous chapter, it may be pointed out here that data collection as well as most of the
front-end processing for HRR is identical to that of SAR. Hence, in this Section, SAR
Figure 2.1 shows the side-looking radar system wherein an aircraft carry on-board
a SAR imager [27-28] illuminates a patch of ground having a target with certain
surroundings. The beam of the radar looks out to the side of the aircraft, in a direction
orthogonal to its flight path. This direction of radiation propagation is referred to as the
Range direction and the direction parallel to flight path is called the Cross-range
energy that impinge on the patch containing the target. Each of these pulses is
performed. The assemblage of data collected and pre-processed in this manner is called a
phase history and is passed on to the processor for image reconstruction. The processor
could either be located on the ground or on board the flying aircraft. This processor gives
out as output the electromagnetic reflectivity of the illuminated ground patch. The
Although a SAR picture looks entirely different from an optical photograph, the
key features are easily recognizable. SAR is coherent radar that employs signal
processing and motion compensation to provide a high spatial resolution estimate of the
scenes reflectivity, also commonly known as Radar Cross Section (RCS). Motion sensors
are used to measure platform flight characteristics so that non-ideal flight path generated
4
phase errors can be removed during image formation processing. Platform or target
scatterers in the antenna footprint. The doppler signatures are subsequently exploited to
4πR/λ. This is the fundamental behind SAR imaging concept (also commonly known as
Range/Doppler imaging).
The reasons for using SAR images over optical ones are summarized below.
• It is able to image a surface with very fine resolution of a few meters to coarse
are chosen.
5
1.1.4: High Range Resolution (HRR)
MTI and SAR are active Doppler systems that transmit and receive electromagnetic
waveforms in the microwave bands that have superior penetrating capabilities than visual
frequency bands. These radar technologies are being researched and developed over
several decades now and both concepts have some share of strengths and weaknesses.
MTI makes use of target movement for image formation and hence, it is highly effective
for distinguishing moving targets from ground clutter. It is a mature radar technology that
allows airborne sensors to survey large areas of land and it has coarse target detection and
range determination capabilities. However, although very useful for target detection, the
MTI technology lacks target recognition capability. In case of SAR, in contrast, ground
target information is available for processing in both range and cross-range domains, and
requirements for SAR is considerably high, preventing it from being used as a wide area
surveillance technology.
Unlike SAR and MTI, the HRR technology considered in this work would rely on
that utilizes SAR image data. Its potential target recognition capability promises to bridge
the gap between the wide area surveillance target detection capabilities of MTI and the
HRR images are used to overcome the disadvantages of SAR data whereas
moving targets are concerned. In case of SAR images, the ability to achieve high Cross-
cells.
6
Figure 1.1: Side looking radar system geometry
resulting in significant blurring due to scattered migration. This becomes evident at low
frequencies since a large coherent processing angle is required for a given Cross-Range
All these factor make recognition hard for moving targets. In case of HRR
profiles, all the information in range is still present, but the cross-range blurring is not
present. This makes HRR as the most feasible choice as far as moving target is concerned
7
1.2: Background and Previous Work
Most research in the field of target recognition address data study, theoretical formulation
and algorithm development. Clearly, important milestones [29-30] have been reached in
these areas. However, barring some notable exceptions [31-33] most existing target
Aperture Radar (SAR) image data. These algorithms are critically dependent on
using SAR is its failure to recognize correctly in case of moving targets due to blurring
unsuccessful in case of moving targets. The other field in which much research is done is
target detection/recognition using Moving Target Indicator (MTI). MTI radar is very
good for detection but fails due to coarse recognition capabilities. In fact, most well
detection/estimation is one of the most important ones [34-36]. An accurate clutter model
had been suggested for precise target detection [37]. The power spectral density (PSD) of
the clutter was estimated such that a multi-dimensional matched filter could be designed
for detection. Another approach [34] has been used for model-based ATR/detection
techniques. The basic paradigm involves detection and feature extraction such that they
can be used in hypothesis using target identities. If the hypothesis is satisfied, the target is
termed as recognized else it is reformed and used to improve the predicted signature.
8
Morgan, et al. [35] has used the Classical Bayesian detection and decision theory for
model-based ATR. It was proposed that when the model tends to represent the
uncertainties in target type, shape, surround, scatterers and feature extraction, then
classical theory yields model based ATR techniques. The concept was extended to use of
model-based templates for SAR-ATR [36]. Mahalanobis [38] has discussed the use of a
Wavelet Decomposition [39-41]. The Wavelet Transform has been found to be highly
effective for image analysis, data and image compression, feature analysis, and many
other applications [42-44]. It has also been used for speckle reduction of SAR images
using a pyramid scheme. Peterson et al. [46] has developed a technique for classifying
Tagaliarini et al. [47] also incorporated the use of Wavelets with Neural Networks. In his
work, the filter coefficients are a linear combination of wavelet coefficients and can give
rise to an energy distribution that makes recognition easier when compared with that of
conventional wavelets.
The use of Eigen vectors corresponding to an Eigen value problem has been
extensively utilized in many applications like Sonar, SAR etc. Bottcher et al. [48] has
presented the optimal method for term expansions based on the optimum eigen function
related to surface of the object. Here, the conversion of Fredholms integral equation of
first kind was done as an eigen value problem of a related hermition operator. This led to
9
target identification by solving the classical scattering theory of waves. Work on ATR
has also been done using Hidden Markov Models (HMM). HMM has been found to be
extremely successful in speech recognition [49] and it has also found some use in SAR
target detection [50]. Liao et al. [8] extracted features from each of the HRR waveforms
wherein the elements of the complex system are implemented as interacting software
objects. New methods have been proposed for use as these software objects. The target
been used in ATR applications to reduce the search combinatorics. These methods use
HRR-ATR has been used to solve the problem of moving target recognition [53-
54]. ATR using HRR profiles has been tried using Neural Networks [55-56]. Yiding et al.
used the property of the distinction of Doppler modulation echo for different targets in
HRR profiles for target recognition. The echo spectral density is obtained by the Fourier
transform. Following that, the choice of the total spectral energy and the four segment
spectral energy as characters is done for use in Neural Networks for ATR. Xun et al. [55]
have used the Matrix Pencil method for scattering centers extraction from full
using Multi-resolution Neural Network. Worrell [56] has used the mean-based templates
for feature extraction. Jacobs et al. [52] has chosen a deterministic Gaussian model for
each Range profile. The likelihood functions under each model for varying orientations
10
and target types are compared. The limit on the orientation estimator performance is
described in terms of Hilbert-Schmidt bound on the estimation error. Stewart et al. [57]
has compared the different classification approaches for HRR profiles. The intrinsic
dimensionality of the signatures was obtained using kth nearest neighbors. The two
classifiers compared were the Gaussian classifier and synthetic discriminant function
(SDF) classifier. In his work, he found that the Gaussian correlation classifier performed
better in presence of white noise while the SDF approach worked better for large angle
bin size.
the target orientation phase behaves to changes in the feature extraction, especially in
For several years, Automatic Target Recognition has been studied for Moving Target
Indicator (MTI) and Synthetic Aperture Radar (SAR) images. MTI and SAR are active
Doppler systems that transmit and receive electromagnetic waveforms in the microwave
bands that have superior penetrating capabilities than visual frequency bands. Though
they are much superiors to optical images they have certain drawbacks when used for
recognition of moving targets. MTI makes use of target movement for image formation
and hence, it is highly effective for distinguishing moving targets from ground clutter but
information is available for processing in both range and cross-range domains, and it has
11
requirements for SAR is considerably high, preventing it from being used as a wide area
degrade when the target is moving because SAR images cannot be formed properly for
Unlike SAR and MTI, the HRR technology would rely on processing High Range
utilizes SAR image data. The information contained in this signature is the radar
scattering characteristics of the target as a function of range along the line of sight of the
radar. It’s potential target recognition capability promises to bridge the gap between the
wide area surveillance target detection capabilities of MTI and the very narrowly focused
target identification capabilities of SAR. Also there is considerable saving in front end
processing in HRR profile generation which require 1-D FFT operation as opposed to
The primary difficulty associated with the HRR sensor for ATR is that it collapses
SAR, making HRR-ATR a more challenging task. Recently Target Detection using HRR
superresolution technique for HRR ATR with High Definition Vector Imaging (HDVI),
where a super-resolution technique is applied to the HRR profiles before the profiles are
Mitchell and Westerkamp [9] for robust HRR radar target identification showed that the
amplitude and location of HRR signature peaks could be used as features for target
classification.
12
Currently, one of the priority research initiatives of the Air Force is to develop an
transition mature HRR-ATR technology into operational Air Force airborne attack and
surveillance platforms. The new HRR-ATR technology can be applied into a system
approach and it is expected to vastly improve Air Force's ability to detect, recognize, as
well as identify time-critical military targets. ATR performance with HRR is found to be
superior for moving targets which cause blurring Synthetic Aperture Radar (SAR) images
harness recent advances from multiple disciplines. At the initial stage, complete
and implementation aspects. This included though not limited to, correlation analysis,
component analysis, signature generation, recognition using Matched Filtering and Least
Squares. Once the interpretation of the basic characteristics of the HRR profiles was
complete, the accumulated insights were eventually gathered systematically in the ATR
performance can be achieved if the training templates are formed via Singular Value
Decomposition (SVD) of detected HRR profiles and the classification is performed using
13
significant proportion (>90%) of target energy is accounted for by the dominant
Eigenvector of the range-space correlation matrix. More interestingly, it was shown that
the range and angle basis spaces are numerically decoupled in the form of left and right
exclusively for the purpose of target recognition. The theoretical results were also
presented to demonstrate that the range space eigenvectors constitute the "optimal"
features in the range domain. Basis space decomposition via SVD is also shown to be
useful for suppression of clutter from measured profile data by eliminating the
Transform when the observation profiles are noisy. Specifically, it was shown that
significant signature information might be lost due to the application of Power Transform
HRR ATR. In [58], the entire 360-degree of a target vehicle circumference was divided
into several optimum-sized sub-targets and templates were constructed from these sub-
targets. Then the result of template matching was combined using Bayesian updating to
60]. But this thesis concentrates on recognition of stationary targets using the MSTAR
data.
14
1.4: Background on Target Detection on SAR images
Recognizing target from SAR images is an important, yet challenging application if the
target is hindered under outliers. To date, the authors have engaged in research and
wideband (UWB) SAR area. The approaches presented include detailed discussion on a
bi-modal technique for modeling ultra-wideband radar clutter was proposed. An approach
to developing a new class of rank order filters, known as, “discontinuity filter” for ultra-
wideband radar target detection applications was presented. These approaches mainly
as well as the development of rank-order filtering algorithms that are designed for basic
UWB SAR sensor phenomenology and at the same time do not require an extensive off-
line training step. Both of these approaches have been shown to generate an acceptable
level of performance under certain conditions that are of interest for UWB SAR
applications.
In this thesis HRR-ATR performance has been analyzed for Moving & Stationary Target
Model (HMM) and Eigen Templates based Normalized Matched filter (ETMF) based
ATR algorithm. The following contributions were made to the existing ATR techniques:
15
A new hybrid 1-D ATR approach is presented where the HRR test profiles are first
scored by ETMF and then the most likely HMM models determined by ETMF are
used for HMM scoring at the second step. Final ATR decision is based on proper
weight combination of the two individual scores. Performance comparison results are
provided for Forced Decision as well as for Unknown Target scenarios. The unknown
target scenario is simulated using the Leave One Out Method (LOOM) [4]. The
In this paper, the proposed hybrid algorithm is extended for moving target case,
which will facilitate simultaneous, multiple target tracking. For Continuous-ID and
joint tracking, the single look ETMF and HMM hybrid technique needs to be applied
observed over time. The proposed approach would be a recursive version of the
performed.
16
approach, a number of categories of algorithms are investigated that implement “self-
training” procedures. These procedures are developed such that a set of localized
regions within a given SAR image are sampled in real-time for purposes of obtaining
low-order and robust real-time clutter models. These real-time models are applied in
a sliding-window type target detection paradigm for clutter cancellation and target
detection. Results are presented from the analysis of three new categories of
algorithms that were developed specifically for this investigation. These three
clutter. Though offline training is required for both EN-SVD and CS-SVD to perform
satisfactory level, the third approach TF-SVD is a notable step to develop a self
training algorithm system i.e. where no offline training is required and the algorithm
A brief overview of the thesis is as follows: Section 2 describes the hybrid approach of
ETMF and HMM. Section 2.1 gives a brief description of ETMF approach; section 2.2
provides a brief overview of HMM training and classification. Section 2.3 explains in
detail the process of combining between ETMF and HMM. Section 2.4 provides the
HMM simulation parameters and also shows the ATR performance results for both
ETMF and HMM individually and the resulting hybrid technique. Section 2.5
17
Section 3 is devoted to explore the performance capability of the proposed time
recursive multiple ID hypothesis. Section 3.1 briefly summarizes the approach and
assumptions, Section 3.2 compares the performance between single profile hypothesis
and time recursive multi profile hypothesis. Section 3.3 summarizes the performance
for UWB radar target detection are investigated. The SAR simulation algorithm is
discussed in detail in section 4.1. A brief discussion of eigen analysis on SAR and clutter
in section 4.3 and comparative detection results are presented in section 4.4 along an
Section 5 presents the conclusion, possible future application and the summary of
this work.
18
2: Robust HRR Radar Target Identification by Hybridization of HMM
sequential HRR radar signatures. The proposed hybrid algorithm combines ETMF and
approach, each HRR test profile is first scored by ETMF which is then followed by
independent HMM scoring. The first ETMF scoring step produces a limited number of
“most likely” models that are target and aspect dependent. These reduced number of
models are then used for improved HMM scoring in the second step. Finally, the
individual scores of ETMF and HMM are combined using Maximal Ratio Combining to
render a classification decision. Classification results are presented for the MSTAR data
algorithm is proposed, where the template features are obtained via Singular Value
performed using normalized Matched Filtering. The SVD operation projects the
information content in a detected HRR profile matrix onto orthogonal basis spaces. This
19
is also known as Karhunen-Loeve Transformation or Principal Component Analysis.
More interestingly, when SVD is applied to a HRR profile matrix, which is range vs.
aspect, it is shown that the range and angle basis spaces are numerically decoupled in the
form of left and right eigen vectors, respectively. This enables us to exploit the decoupled
eigen vectors constitute the "optimal" features in the range domain. In addition, SVD
analysis of a large class of MSTAR targets indicates [12] that over 95% of target energy
is accounted for the largest singular value only, further justifying the proposed utilization
SVD is also shown to be useful for suppression of clutter from measured profile data by
Most work on Automatic Target Recognition (ATR) has been performed using Synthetic
Aperture Radar (SAR) images. ATR using SAR images performs poorly in case of
moving targets due to blurring caused in the cross-range domain. The HRR-ATR
traditional SAR-ATR that utilizes SAR image data. In HRR based ATR systems there is
only 1-D FFT operation, as opposed to SAR's use of 2-D FFT. The processing factor
20
SAR image, radar returns must be generated over a relatively large sector of angles. With
HRR profiles, only a relatively small number of angles would be sufficient to perform
ATR. Figure 2.1 shows the process of generating HRR profiles from Complex Phase
History (CPH). As shown, SAR image can be obtained from the HRR profiles by taking
depression angle of radar. This makes a fixed sampling window. The two primary HRR
waveforms for SAR systems are the Frequency stepped and Linear Frequency
modulation. The Range resolution (∆R) is determined by the radar RF bandwidth. Thus,
N 2
j4 πR i
Y(τ j ) ∝ ∑
i =1
σi e λ
(2.1)
Where σi is the RCS of elemental scatterers in Range gate, Ri is the Range and N is the
Complex phase
IFFT (r2+x2)1/2 SVD
history (CPH) Complex Detected Eigen
HRR HRR profiles Templates
21
Note that no power transform operation is performed on magnitude HRR as in [12] we
proved that, power transform severely degrade ATR performance if noise is embedded
2.1.2: Normalization
The template profiles of all the targets are normalized to have same length (i.e. energy),
while preserving their angular separation and relative variations in scattering returns.
The HRR profiles of the Segmented Data set provided by AFRL (TRUMPETS) are not
aligned in Range. Hence each Profile of 1-Degree Sector should be aligned with respect
to each other. This alignment is achieved by taking a profile as a reference and shifting
the adjacent profile till maximum correlation was achieved. This procedure is repeated
until all the profiles in a sector have been aligned. Though this procedure of aligning the
Singular Value Decomposition (SVD) is a very effective and robust tool for decomposing
any matrix into orthogonal basis spaces. Let Y be an NXM of detected range profiles at
M angular looks containing N range gates each. The SVD operation produces basis
• Y ∈ ℜ NXM: Detected HRR Profile Matrix, N = No. of Range profiles and M = No.
of Angular looks
22
M
Y
→ UΛV =
SVD T
∑λ u v
i =1
i i
T
i (2.2)
where,
• Singular Values :
e
d
ut 3
i
n
g
a
M
2
0
0 5 10 15 20 25 30 35
Number of Eigen Values
Fig. 2.2: Distribution of Singular values for MSTAR target T72, 1000 sector
23
Where, EV[.] denotes the operation “Eigen-Vectors of”. For Range vs. Angle
HRR data, the left eigen vectors (U) span the orthogonal basis space in the range domain
while the right eigen vectors (V) span the angle space. The middle matrix Λ is diagonal
λ11 ≥ λ 22 ... ≥ λ MM , where λ ii denotes the weights associated with i-th eigenvector. Larger
the target signal. Hence these are denoted as “signal subspace” eigenvectors whereas
those corresponding to the smaller singular values are denoted as “noise or clutter
subspace”. Figure 2.2 displays the distribution of singular values for a typical MSTAR
targets in a particular degree range. In that case, it is seen that only the highest singular
value ( λ11 ) makes up more than 96% of the total energy of the distribution. Interestingly
the range space in U and the angle space eigenvectors in V appears in decoupled form
after the SVD operation is applied to Y as shown in equation (2.3). It can be concluded,
the HRR profile matrices are close to rank one, which implies that u 1 , the left-
contain the essential range information of the underlying target. Hence, here it is
proposed to use the dominant range-space (left) eigenvector as the feature template.
24
2.1.5: Unknown Target Classification
Template
|<--------------------------Overlap Region-------------------------->|
Test Profile (shift =+ 8)
Given observed (or, test) range profile(s) of an unknown target, the ultimate
accomplished by comparing the observed profile with all the available templates, which
are assumed to have been formed beforehand using training data set. The decision
determines the target type for which the correlation between its template ( mi ) and the
observation (a) profile is maximized among all template choices. However as the
observation profile a and all the template may not be exactly aligned, the correlations
have to be calculated with various lag values and the maximum correlation among all
lags for each target type has to be determined. For each target, there are usually a large
25
number of templates at different aspects. In our simulations with MSTAR database,
The maximum correlation value among all templates within ± D° of target aspect
(assumed known or estimated by an MTI tracker) for each target is determined. This
process is repeated for all target classes, with each class being assigned its maximum
correlation out of all lags for aspect angles within ± D° . Finally, the target class having
the maximum correlation value among all classes is termed the matched target class. In
our simulations, correlation lag values up to ±5o of the true aspect was used because it is
assumed that the MTI tracker (running in conjunction with HRR-mode radar) would
To improve the performance of the ATR algorithm it is important to include that portion
of the Observation and Template profiles which contains significant portion of the target
signature information. Therefore, if the Observation and Template profiles are not pre-
aligned it is important that they be aligned prior to using them in the classifier. In this
work the Centroid of a range profile was used as the reference in aligning the
As described in the previous section, the Matched Filter Classifier assumes that
both the Observation and the Template profiles are normalized to have equal lengths.
However, while correlating the template and test profiles to find the best match, one of
the profile vectors is shifted to the left and right of the Centroid to obtain the maximum
correlation. When the observation profile is shifted over the Template profile, the region
26
of overlap between the two would change with each shift. However, the norms of the
overlapped regions of the observation and template may also change with each particular
shift. Hence, using a stored template profile originally normalized over its entire length
will not be appropriate if used as is. In order to satisfy MF’s requirement that both
template and observation have identical lengths, it is important that only the overlapping
parts of both the profiles is normalized prior to correlating the vectors, as described next.
Let the test and template profiles be represented by narrow (shaded) and wide
rectangles, respectively, as depicted in Figure 2.4. The lengths are shown different
intentionally, as the test and template could be of different lengths. Different heights are
used primarily to differentiate between the test and template. It has no other implication.
A
OBSERVATION PROFILE Template Profile
Fig. 2.4: Observation and Template Profiles are shown in shaded and blank boxes of different lengths
Next for better understanding, the correlation process with overlap normalization
is described in detail. The test profile was shifted over the template and correlated. In the
next figure, it is assumed that the shift is –8 with respect to the centroid. Clearly, the
entire lengths of neither test nor the template are overlapping. Hence, it doesn’t make any
sense to normalize over entire lengths of the template or test, because the correlation is
27
occurring only over the overlapped (shown in stripe) region. It will be more appropriate
to ensure that the norms within the overlap region of the vectors are kept the same.
Hence, we re-normalize both vectors only over the overlapped parts (in stripe) before we
perform correlation.
Next, the case when both test and templates are aligned on the Centroid is
depicted. In this case, the entire length of the template is overlapping some middle
portion of the test. Hence, once again, we re-normalize only within the overlapping
regions to ensure that both vectors have same lengths. It may be noticed that the length of
the overlapped portion (in stripe) of the vectors is longer than the previous case.
28
Overlap region to be normalized
OVERLAP REGION
Next, the +8 shift case from centroid is shown. Again, the overlapped regions
have changed for both. Again, only the striped regions are normalized for both vectors
before correlating.
29
2.2: HMM approach of training and classification
Some recent work has shown encouraging promise for the use of HMM in HRR target
recognition [8]. The airborne radar transmits microwave pulses at constant depression
angle. Each pulse is reflected from the target and gets back to the radar receiver.
sequence of range scattered pulses which is termed as High Range Resolution (HRR)
profiles. The HRR signatures characterize the target at a specific airborne sensor
orientation. In the MSTAR data collection studies, it is assumed that the depression angle
of airborne radar with respect to the target is constant and the target sensor orientation is
HRR signatures at different orientations, the scattered range field can be assumed to be
stationary over small angular sectors. Each such angular region is termed as a “state”. As
the target orientations are unknown in addition to target identity, theses information can
be assumed as hidden and HMM can be used to model and characterize the sequence of
scattered waveforms. Figure 2.8 shows a framework for HRR-based target recognition
using HMM.
HRR profiles
(Training data) Training Model
(HMM)
Code Book
30
This subsection describes an introduction to Hidden Markov Models and algorithms for
transition from one state to another. Unlike a classical Markov model, a Hidden Markov
Model introduces an output probability density function (Pdf) to define the conditional
probability that a symbol is generated from a finite set of symbols, given that we are in a
process with an underlying Markovian finite-state structure that may only be observed
indirectly (hence the “hidden” nature of the model). At any given time, it is unknown to
an outside observer what state the process is in, but it can be observed through the
We will limit our consideration to the first-order Hidden Markov Model, where
state dependencies are on the immediate predecessor only. Another assumption made in
depends only on the current state at this observation frame and is conditionally
independent from the previous state and the past symbols emission. This second
assumption may degrade the experimental realism of HMM’s, but it reduces the number
of parameters required by the model and allows the use of efficient evaluation and
31
2.2.1.1: Elements of HMM
Fig. 2.9: Two states Hidden Markov Model with two output symbols, V1 and V2
Figure 2.9 shows a Hidden Markov Model with two output symbols, V1 and V2. This
simple model is used to explain the elements of HMM. The parameters of the HMM that
can generate the output symbols V1 and V2 are shown in Equation (2.6) and (2.7).
N = 2, M = 2 (2.6)
N: the number of states in the model. We will denote the individual state as
32
M: the number of distinct observable symbols per state or the size of the codebook.
A: NxN matrix representing the state transition probabilities, i.e. the probability to
And
N
a ij ≥ 0, ∑a
j=1
ij = 1 ,1 ≤ a ij ≤ N (2.9)
B: NxM matrix which specifies the observation symbol probability distribution in the
state sj:
And
M
b j (k) ≥ 0, ∑ b (k) = 1 ,
k =1
j 1≤ j≤ N (2.11)
The complete parameter set λ of HMM requires the specification of two model
parameters (N and M), the specification of the observation symbols, and the specification
33
of the three probability matrices A, B and π. The compact but convenient notation is used
λ = ( A, B , π ) (2.13)
The use of HMM models in real-world applications requires the solution of the following
efficiently evaluate the probability that the model generated the observations?
determine an optimal state sequence in the model that generated the observations?
Training: Given a model and a set of observations, how do we adjust the model
In the following sections, the solutions proposed for each of these three basic problems
are reported.
The evaluation problem can be stated as: given the observation sequence
that the observed sequence is produced by the model. The most straightforward way to
compute this is to enumerate all possible paths (state sequences) of length T that generate
34
P ( O | λ ) = ∑ P ( Q | λ ) P ( O | Q, λ ) (2.14)
all Q
Where Q is the state sequence: Q = q1q 2 L q T and q1 is the initial state. The first factor in
T
P ( Q | λ ) = ∏ P(q t | q t −1 ) = πq1 a q1q 2 a q 2q 3 L a q T−1qT (2.15)
t =1
The second factor in Equation (2.14) can be re-written by applying the output-
independence assumption:
P ( O | Q, λ ) = bq1 ( O1 ) ⋅ b q 2 ( O 2 ) L bq T ( OT ) (2.16)
P ( O | λ ) = ∑ P ( Q | λ )P ( O | Q, λ ) (2.17)
all Q
= ∑
q1 ,q 2 ,...q T
πq1 b q1 (O1 ) a q1q2 b q2 (O 2 )...a qT−1qT b qT (O T ) (2.18)
From Equation (2.18) we can directly calculate the P (O λ ) from the HMM parameters,
but the computation is unfeasible even for small values of N and T because the
to compute P (O λ ) recursively.
35
2.2.1.4: Optimal State Sequence
Given a model λ and a sequence of observations O = O1O2 ...OT , one problem that needs
to be addressed is the estimate of the best state sequence Q = q1q 2 ...q T (or the most likely
method called Viterbi algorithm [60] is used to choose the optimal state sequence, i.e., to
The training problem involves adjusting the model parameters in order to maximize the
probability of the training observation sequences being produced by the model. The
iterative procedure called the Baum-Welch algorithm is used to choose the maximum
likelihood model parameter λ such that its likelihood function P(O | λ ) is locally
maximized.
2.2.2.1: Framing
Here, each HRR profile is blocked into frames of N samples, with adjacent frames being
separated by M (M<N). The first frame consists of the first N samples. The second frame
begins Mframe samples after the first frame, and overlaps it by N- M samples. Similarly,
the third frame begins 2M samples after the first frame (or M samples after the second
frame) and overlaps it by N- 2M samples. This process continues until all the profile is
accounted for within one or more frames. As each HRR profile consists of a small
36
number of samples, framing is pretty useful to duplicate the information contained in a
single profile and learning the HMM. Framing is done on HRR profiles used both for
2.2.2.2: Clustering
The next specific step is to make target and aspect specific Vector Quantization (VQ)
commonly used LBG algorithm [61] for clustering a set of L frame vectors into a set of C
1) Design a 1-vector codebook; this is the centroid of the entire set of training vectors
2) Double the size of the codebook by splitting each current codebook yn according to the
rule:
y +n = y n (1 + ε) (2.19)
y −n = y n (1 − ε) (2.20)
Where n varies from 1 to the current size of the codebook, and ε is a splitting parameter
3) Nearest-Neighbor Search: for each training vector, find the codeword in the current
codebook that is closest (in terms of similarity measurement, here we have taken as
Euclidean distance as similarity measure), and assigns that vector to the corresponding
4) Centroid Update: update the codeword in each cell using the centroid of the training
37
Iteration 1: repeat steps 3 and 4 until the average distance falls below a preset threshold.
Figure 2.10 shows, in a flow diagram, the detailed steps of the LBG algorithm.
“Cluster vectors” is the nearest-neighbor search procedure that assigns each training
vector to a cluster associated with the closest codeword. “Find centroids” is the centroid
update procedure. “Compute D (distortion)” sums the distances of all training vectors in
Find
centroid
Yes No
m<M Stop
Split each
centroid
m = 2*m
Cluster
vectors
Find
centroids
Compute D
(distortion)
No D'−D Yes
D’ = D <ε
D
38
Intuitively, the LBG algorithm designs a C-vector codebook in stages. It starts first by
initialize the search for a 2-vector codebook, and continues the splitting process until the
desired C-vector codebook is obtained. Clustering is done on HRR frames used for both
Euclidean distance is computed for each HRR frame with all codebook centers.
The codebook centers, which are the nearest of the frames, emerge as winner and the
indices of those winner centers are sent to HMM optimum parameter estimation process
during training mode and maximum likelihood computation process during testing mode.
HMM model parameter estimation i.e. estimating parameter λ such that its
parametric model.
HMM Optimal State Sequence estimation i.e. given a model λ and a sequence of
of the best state sequence Q = q1q 2 L q T (or the most likely state path) corresponding
39
2.2.2.3.1: Model Optimum Parameter estimation:
In the proposed HMM training, one HMM model is designed for each target of interest.
Each target HMM model is subdivided into its corresponding aspects. A target model is
partitioned into N distinct states, denoted by the set S={ s1 , s2 ,...,s N }, where the total
number of states N can be determined from azimuthal partitioning. A HMM model can
where, aij = P (st+1 = p| st = q), the probability of being in state p at time t+1, given that at
time t the state is at q. The state output probability B={bij}, bij = P(uk at t| st = q), where
U={ u1 , u 2 ,..., u N } is the observation likelihood sequence from all N states at time t. State
In the ATR application, these initializations are dependent on the angular extent of each
state and the difference between target orientations over successive measurements. If θi
represents angular extent of state i and ∂θ the change of target orientations over 2
consecutive measurements, the initial state transition probabilities can be defined as:
40
As we have no prior information about the targets, the initial state probability for each
θ j,i
πj = N
(2.24)
∑ θ j,i
i =1
where the subscripts i and j are the state and target indices respectively.
If at a particular time instant t, Ot is the trained HRR profile, our primary task in the
the likelihood P (Ot| λ ) is maximized. The optimal A, B and π matrices for each target
can be estimated using the well-known recursive method called Baum-Welch re-
estimation algorithm [62]. This is a Bayesian approach where an initial HMM model is
First, state occupation probability at time t given the observation sequence and
γ t (i) = P(q t = s i | O , λ ) i.e. the probability of being in state i at time t, given the target
observation id O = O1, O2…OT and the model λ . qt is the state observed at t-th instant
41
Where, α t (i) is the forward variable and can be defined as:
α t (i) = P (O1, O2… Ot, qt = si| λ ) i.e. is the probability of partial observation sequence up
to time t and the state si at time t, given the HMM model λ . (2.26)
time t+1 to T, given the state si at time t and the model λ . (2.27)
making a transition to state j at time t+1, given the observation target id O = O1,O2,…,OT
P(q t = s i , q t +1 = s j , O | λ )
ξ t (i, j) = (2.28)
P(O | λ )
But O can be defined as O = O1, O2… OT. Hence the Numerator can be written as,
= Term1×Term2 (2.29)
From Term2 it is evident that we have to choose the observation id Ot+1 from state sj.
Hence we have,
= aijbj(Ot+1) β t +1 ( j) (2.31)
42
α t (i)a ijb j (O t +1 )β t +1 ( j)
ξt (i, j) = (2.32)
P(O | λ )
Fig. 2.11: Illustration of the sequence of operation required for the computation of the joint event that the
Figure 2.11 shows the computation of ξt (i, j) based on the forward and backward
variables.
From equation 2.32, α t (i) accounts for O1…OT, aij accounts for transition
probability from state i to state j, b j (O t +1 ) picks up the symbol O t +1 from state j and
The expected number of transitions from state si during the time instant t is:
T −1
∑ γ t (i) (2.33)
t =1
43
The expected number of transitions from state si to state sj during the time instant t is:
T −1
∑ ξt (i, j) (2.34)
t =1
πˆ i = γ t (i) ,1 ≤ i ≤ N (2.35)
T −1
∑ ξ t (i, j)
â ij = t =1
T −1
,1 ≤ i ≤ N (2.36)
∑ γ t (i)
t =1
T
∑ γ t ( j)
t =1
Ot = k
b̂ j (k) = T
,1 ≤ i ≤ N (2.37)
∑ γ t ( j)
t =1
44
The EM algorithm itirates by using the new HMM parameters λ computed in (2.35)-
λ converges. The HMM parameters can be iteratively computed as long as P (O| λ t +1 ) >
P (O| λ t ). When the inequality becomes equal or if the right hand side becomes less than
the left, we can conclude that λ is converged and the estimated λ̂ maximizes P(O| λ )
i.e. upon convergence of λ , we can conclude that the likelihood is maximum of a given
model.
estimate the best state sequence Q = q1q 2 L q T corresponding to the given observations.
A famous dynamic programming method called Viterbi algorithm [60] was used to
maximizing P (Q, O λ ) . We define the maximum probability along a single best path at
time t, which accounts for the first t observations and ends in state sj given the HMM
model λ, as
We also define an array ψ t ( j) to keep track of the argument, which maximizes δt −1 (i)a ij
for each t and j. The Viterbi algorithm obtains a solution recursively as follows:
45
Initialization:
ψ1 (i) = 0 (2.40)
Recursion:
Termination:
Path Backtracking:
q ∗t = ψ t +1 (q *t +1 ), t = T − 1, T − 2, L , 1 (2.45)
From the above induction, we can see that the Viterbi searching is a time synchronous
searching algorithm. Before going to time t+1, the process at time t is completed. For
time t, each state is updated by the best score from the states at time t-1. After finishing
the forward state estimation, Viterbi algorithm backtracks to optimize the states for the
46
2.2.2.4: HMM Classification
λ1 HMM Model 1
Likelihood
computation
λ2
Observation
sequence Collection of
Likelihood
Computation likelihood’s
Framing Vector (P(O|λ)) indicating
Quantization the similarity
Test between the test Target index
HRR HMM model K HRR profiles and corresponding to
Feature the (PHMM)max
vectors training models
λK
Likelihood
Computation
In the training procedure, one HMM model is optimized for each target of interest. In the
testing process, a given set of sequential HRR profiles under test are submitted to all
HMM models, and the data is associated with that target for which the respective HMM
yields the maximum likelihood, i.e. for each test HRR profile at time t, we computed
P(Ot| λ k ), where k, the index, varies from 1 to the total number of HMM models. The test
profile will be assigned to that target class whose corresponding HMM models produces
the largest likelihood. The classification procedure can be understood from the block
diagram 2.13.
47
The HMM classification problem can be mathematically stated as: given the
varies from 1 to total number of HMM models (which is same as number of trained
targets), compute likelihood P ( O t λ k ) for each HMM model and classify the unknown
observation profile at time t to that known target class, whose likelihood is the largest.
recursively for each HMM model and for profiles observed at each time instant. The
i.e., the probability of the partial observation sequence, O1O 2 L O t , (until time t) and
The forward and backward recursion procedures are depicted in the next page.
48
Fig. 2.14: Illustration of the sequence of operations required for the computation of the (a) forward variable
49
We can compute α t (i) inductively, as follows:
Initialization: The initial forward variable α1 (i) is the joint probability of the state i,
It can be seen that state j can be reached at time t+1 from the N possible states i,
α t +1 ( j) = P ( O1O2 L O t +1 , q t +1 = s j | λ ) (2.48)
N
= ∑ α t (i)a ij b j (O t +1 ), where 1 ≤ t ≤ T − 1, 1 ≤ i, j ≤ N (2.49)
i =1
Termination: The probability of the complete observations sequence is the sum of all
N N
P ( O | λ ) = ∑ P ( O1O2 L O T , q T = si | λ ) = ∑ P ( O, q T = si | λ ) (2.50)
i =1 i =1
N
= ∑ αT (i) (2.51)
i =1
The forward probability calculation is based on the lattice or trellis structure showed in
Figure 2.15. Since there are only N states, all possible state sequence will remerge into
50
S S S S
α(y1t −1 , j)
β(y Tt+ 2 | j)
j j j j
α(y1t ,i)
β(y Tt+1 | i)
a(i | j) i a( j | i)
2 2 2 2
b(y t +1 | j)
b(y t | i)
1 1
1 1
t−2 t −1 t t +1 t+2
yt−2 y t −1 y t y t +1 y t+ 2
In Figure 2.15 above, the time instants are denoted as t and the observation sequence at a
51
2.2.2.4.2: The Backward Procedure
Figure 2.14 (b) illustrates the backward procedure. We define a backward variable:
i.e., the probability of the partial observation sequence from t+1 to the end, given state si
at time t and the model λ. βt (i) was computed using the following recursive steps:
Initialization:
Induction:
N
= ∑ a ijb j (O t +1 )β t +1 ( j), where t = T − 1, T − 2, L ,1, 1 ≤ i ≤ N (2.55)
j=1
Termination: The probability of the complete observations sequence is the sum of all
N N
P ( O | λ ) = ∑ P ( O T O T −1 L O1 , q1 = s i | λ ) = ∑ P ( O , q1 = s i | λ ) (2.56)
i =1 i =1
N
= ∑ β1 (i) (2.57)
i =1
algorithm to compute the probability of the observation sequence given the model. The
classification algorithm will assign the observation sequence O to that target class whose
52
2.3: Approach of Combination between ETMF and HMM
As, single ETMF or single HMM failed to provide satisfactory target recognition
performance (shown in later section), concentration was directed to hybrid those single
Each ATR algorithm has inherent performance limitation. ETMF or HMM cannot
exceed it’s maximum capability and perform better than that. Hence single algorithm
ATR classification is not always the correct way to get above satisfactory level
output.
The major advantage of the hybrid technique is: if one algorithm fails to get target ID
correctly, it is quite possible that another algorithm might work. In the hybrid
probability of having mistake for both techniques at the same time/same place will
reduce drastically. The challenge is to design a hybrid system that can recognize the
target correctly even if only one algorithm between the two makes a correct decision.
In other words, the main question to be answered during the design phase is how to
improvement.
53
2.3.2: Proposed Solution
There were hybrid algorithms in the literature [63] applied in speech recognition, where
the same set of data and produce discriminant scores independently. Next, the
discriminant scores produced by those approaches were combined with proper weights,
which are derived from the data set, and classification was performed on the combined
set of discriminant results. The primary advantage of the hybrid algorithm is improved
classification performance can be achieved compared to each algorithm alone, but the
classify unknown target is a major area of concern. Considering the computational need
and performance improvement, in this paper, the hybrid algorithm presented is modified
version of what being already seen in the literature. In this proposed hybrid algorithm the
two classification algorithms are not allowed to run parallely, in fact the operation of two
ATR algorithms are rather serial and the operation of the second algorithm is much
dependent on the successful act of the first. The major steps of the proposed algorithm are
described below:
In the first step, one ATR algorithm is allowed to choose several Most Likely training
models, which can produce correct classification. Each most likely models are
assigned discriminant scores. The discriminant scores attached with each training
model indicate how closely that model matches with the unknown observation
sequence.
54
Hence the task of first ATR algorithm can be diagonized as :
Assigning a discriminant score with each one at the smaller subset indicative of
how closely the models match with the unknown observation sequence.
Next, smaller subset model information is passed to the other ATR algorithm.
The task of the second ATR algorithm is to perform classification of the observation
waveforms with only those smaller subset of models and also assign discriminant
scores with each member of the smaller subset group indicating how closely the
match is between the known trained model and the unknown observation sequence.
Hence for a single observation sequence we have two sets of discriminants scores
Next, the two sets of discriminant scores are combined with weights estimated from
the data and accuracy of one ATR algorithm over the other and finally the unknown
observation sequence is decided as that of the known trained target whose model
In unknown target scenario, the highest combined discriminant score achieved for the
“unknown”.
In the proposed hybrid approach, ETMF and HMM techniques perform the functionality
as first and second ATR algorithms respectively. The necessity and usefulness of
55
2.3.2.1: Necessity of developing modified hybridization in ETMF ATR
In the ETMF testing phase, each HRR profile is tested with 3-aspect angle templates
trained target templates. Here we are assuming that MTI radar is also in operation and
that is in synchronous with HRR radar. As MTI radar is pretty accurate in calculating the
model searching for each target, when an unknown observation sequence comes in. The
ETMF algorithm chooses the target corresponding to the highest discriminant correlation
value. However, it will take wrong decisions in some cases leading to False Alarm or
Missed target. One effective way to reduce the probability of wrong decision is to couple
the ETMF technique with some other independent ATR technique(s) and then make the
final target classification decision based on a combined result. The inherent assumption
here is that the probability of failure in classifying a target by both ATR techniques at the
same time would be very low if the target classification methods are independent of each
other. Since HMM is now well established in speech and pattern recognition, the ETMF
results are combined with that of HMM and the classification decision is taken based on
the combined performance. But here is the problem, incorporating HMM to classify
unknown targets has a huge drawback. As during testing, HMM has to go through
number of target templates are large, it goes beyond real time. Hence the tricky issue was
That is the reason why we modified the hybrid approach and instead proposed a
56
selects a small number of most likely models that are most similar to the unknown
observation sequence. In the next stage, HMM likelihood computation is limited to those
numbers of models and it computes discriminant scores for those models only. By doing
this, a lot of computation time was saved as HMM need not perform matching with all
it’s models between ± 1o . Here it is assumed that ETMF is accurate enough to pass those
models where the actual solution is embedded to HMM. This assumption is based on
that though sometimes it misses the correct target and assigns highest discriminant score
to the wrong one, but the correct target was always among the top few output
discriminants. Hence we decided to use only a subset of the higher discriminants values
for HMM to score. This method not only saves computational efficiency and implements
the hybrid system in real time but also provides better recognition rate compared to the
The block diagram 2.16 depicts the proposed combining procedure between ETMF and
HMM.
57
The detailed description about number of subset model selection and weight calculations
to combine ETMF and HMM output discriminant scores are provided in the later section.
The proposed hybridization technique is described in detail in the last two sections. In a
recent hybridization technique in literature, all the HMM models are used in scoring [63].
In our proposed hybrid technique, only few number of HMM models that can most likely
generate the correct solution are used in discriminant scoring. The specific models that
HMM will use in scoring are to be determined by the ETMF algorithm. As HMM has
models for each target and aspect (same as ETMF), only those HMM models are used to
score the observed HRR profiles that give high discriminant scores in ETMF testing. It is
important to emphasize that the exact number of HMM models to be used for scoring is
determined from the training HRR profile data. To find out the optimum models that
HMM will use in classification, the training data was split into training and testing sets.
The “testing set” was formed by randomly selecting 25% of the HRR profiles from each
target in the training database. The remaining 75% of HRR profiles from each target are
regarded as training set and are used to generate the Eigen-templates and HMM models.
At each aspect the “test” set was then used to calculate the HMM classification
HMM model goes accordingly to the maximum correlation score assigned by ETMF on
58
each model. In other words, at every aspect, classification for the same aspect test
waveforms are performed using HMM models from 1 to 3K (K is the number of targets
in the database) successively according to the ETMF discriminant scores. The target-
aspect model that is assigned highest similarity by ETMF is used first on HMM
classification. Then the same aspect all test data is classified again by the top two
discriminant target aspect model. This process will go on until all 3K HMM models are
used in classifications. In order the system not to be biased for the specific test samples,
ten different Monte Carlo simulations are performed by taking ten random collections of
“testing set”. The % recognition rate shown in the Figure 5 is the average of all ten HMM
recognition results. The number of HMM models those gives the highest average
recognition rate in that aspect is saved and is used during real testing. The same process
is repeated for all aspects from 10 to 3600. Hence HMM model selection is an aspect
dependent process. Figure 2.17 shows a case of optimal model selection for HMM
classification.
59
Fig. 2.17: In 100 aspect case, this plot shows the HMM recognition rate with number of HMM model
(In our simulation as we used 4 targets, maximum no.of HMM model is possible to score is 12)
From Figure 2.17, it can be seen that for train HRR profiles in the 100 aspect case, HMM
recognition rate is maximum when we used 3 most likely models, assigned highest
similarity by ETMF. Hence, 100 aspect unknown observational profiles are classified
using the 3 HMM model which would get highest ranking in the preceeding ETMF step.
This technique of reduced HMM model scoring provides a 10-15% improvement in PCC
as compared to the case when all the HMM models are applied for scoring purposes.
60
2.3.2.2.2: Weight determination
In order to facilitate the hybridization with HMM, the correlation based ETMF scoring
correlation is small, then the probability that the profile is the desired target will also be
small, and as the correlation increases, the probability increases and approaches one.
Assuming ETMF and HMM techniques are statistically independent, the joint probability
where PCombined, PETMF and PHMM are discriminant likelihood’s indicating similarity
between the train templates and the unknown observation obtained from the hybrid
indicates the discriminant score achieved from the hybrid combiner is equal to the
multiplication of that obtained individually from ETMF and HMM. This condition is
achievable because ETMF and HMM similarity measures are independent of each other.
This would be the simplest combination technique where the weightings on both ETMF
and HMM are equal. However, since the performance of ETMF in classifying targets is
may be achieved by feeding longer sequence of HRR profiles for estimating the HMM
training model parameters [62], which may not preferable for moving targets), the
61
proposed hybridization approach applies different weights to the log discriminant scores
The weights W1 and W2 are determined from the training data and they are target and
aspect independent. For the 4-targets MSTAR data used in our simulation, the best choice
for weight proportions is 75% and 25% for ETMF and HMM, respectively. This decision
was made from Figure 2.18 (Combined recognition rate vs. W1/ W2).
Fig. 2.18:This figure is used to determine the most effective W1/ W2 so that the combined
ETMF+HMM recognition rate is the highest.
62
To find out the optimum W1/ W2, the training data was again split like optimum model
selection into training and testing sets. The “testing set” was formed by randomly
selecting 25% of the HRR profiles from each target in the training database. The
remaining 75% of HRR profiles from each target are regarded as training set and are used
to generate the Eigen-templates and HMM models. The “test” set was then used to find
the combined classification performance by varying W1/ W2 (as seen from the Figure
2.18). In order the system not to be biased for the specific test samples, ten different
Monte Carlo simulations are performed by taking ten random collections of “testing set”.
The % combined recognition rate shown in Figure 2.18 is the average of all ten combined
W1 3
recognition rates. It is found out that for = , the combined ETMF and HMM
W2 1
produced the maximum recognition rate. Hence, the same 3:1 ratio of W1/ W2 was later
used in actual test HRR profile scoring. Hence the final ETMF+HMM combined
which ensures that the summation of W1 and W2 is equal to 1 i.e. W1+W2=1. The hybrid
ATR algorithm selects the highest logarithmic combined score and classifies the target
accordingly. In the simulation section it is shown that the combined ETMF-HMM hybrid
63
2.4: Simulation Results
The proposed algorithm was trained and tested using MSTAR data set (without power
transform) containing HRR profile sets of 4 ground military vehicles (BMP2, T72, 2S1
and BRDM2) at 170 depression angle over 360 degree of aspect angles. The MSTAR data
set typically consists of SAR image chips. These images have been converted to the HRR
profile sequences through 1-D inverse Fourier Transform in the cross-range domain and
several filtering operations. The raw HRR profiles used in these studies were formed
from SAR imagery under the Defense Advanced Research Projects Agency (DARPA)
TRUMPETS Program and provided to Wright State University by Air Force Research
ETMF based templates are generated from the MSTAR 170 depression training data set.
The templates at each target-aspect are formed by computing the eigen vector ( u 1 )
corresponding to the highest eigen value of the range space correlation matrix R̂ , where
ˆ = YYT , where Y={r1, r2,…, rN} i.e. the range profile matrices at each aspect.
R
64
2.4.2.2: Classification using Matched Filter Technique
In the ETMF technique, the observational profile a is correlated along ±8 shift with the
template profiles generated from the training HRR dataset. Before correlation, two
important preprocessing steps are performed which ultimately enhances the recognition
rate. The test profile is centroid aligned with respect to the templates and the overlap
region between test and templates are normalized prior to performing correlation. In the
Matched filter based classifier, each test profile is correlated against ± 10 HRR templates.
2.4.3.1: Framing
Each HRR profile is blocked into frames of N samples, with adjacent frames being
separated by M (M<N). In the HRR framing part, N=12 and M=8 were selected.
2.4.3.2: Clustering
Vector Quantization (VQ) rule is followed for clustering the HRR frames. The commonly
used LBG algorithm for clustering a set of L frame vectors into a set of C codebook
vectors was formally implemented. Clustering is performed with cluster center (C)=32
(for this data C=32 was found to be a reasonable trade-off between performance and
computation efficiency). The clusters are target and aspect dependent. As MSTAR
dataset has K=4 targets with aspect angle data from 00 to 3600, we have 360 × 4 cluster
center matrix. To make the clusters more robust, HRR frames of a particular degree are
65
mixed with adjacent degree HRR frames before clustering and clustering is performed on
the combined HRR frame matrices, i.e. a particular degree cluster not only consists
profile information for that particular degree but also it’s nearest neighbors. For example,
the 20 cluster will have information not only for 20 but also for 10 and 30, and so on.
In the next step of this classification algorithm, one HMM is designed for each target, i.e.,
there is a total of K HMM models, where K=4 in our simulation. Another important
factor for HMM design is to select the number of hidden states. The total training HRR
profiles are divided into N states using a uniform angular decomposition (for N states,
each state has an angular extent of 3600/N). In this algorithm, N=120 was used, so that
In this thesis the result of the proposed ETMF–HMM hybrid algorithm is compared with
those of the individual techniques in single look case. The results are presented for three
different classifier approaches, Hybrid classifier, ETMF based classifier and HMM based
classifier. The classifiers are designed in such a way that they can reject unknown targets
while maintaining high known target classification decision. The MSTAR data set used
66
shown Power transform degrades the classifier performance severely for noisy scattered
profiles.
the hybrid classifier over that of it’s component algorithms. Simulation results are
presented in the form of confusion matrices and are given in tables II-XI. In these tables,
the performance results for a single independent look are presented. Multilook techniques
can certainly improve the recognition rate further [6-7]. But as conventional Multilook
techniques are difficult to implement in tracking, results are presented in a new time
recursive multi target hypothesis updating procedure, which is time recursive version of
the stationary multilook approach and tracker friendly. Two types of classification tests
are performed, namely Forced decision and unknown target scenario. The former
assumes that all test targets belong to one of the known training target classes, which is
the unlikely in practical situation. The later makes no such assumption, so each test target
needs to be compared with a threshold before making any decision whether it is known or
unknown.
In the Forced Decision case, the classification of the target class with the largest
likelihood determines decision for an observation profile. Table II summarizes the Forced
decision result for the above mentioned three algorithm. In table II, PCC comparison is
made among the proposed hybrid technique, ETMF alone and HMM alone for single
look-single profile, single look-3 profile average and single look-5 profile average
measures. It is found there, PCC of the ETMF-HMM hybrid technique is the highest for
67
all three cases. The PCC for ETMF, HMM and hybrid displayed in table I is actually the
average of all four target’s individual PCC’s. It can be also seen that the performance of
the HMM-only case is relatively poor, but when it is combined with the ETMF technique,
observed from table II. However, for moving targets and tracking applications the
position of target changes with time. Hence, it is advantageous to use a single profile at a
time to obtain instantaneous track information. This is the reason for us to show the
classifier result on single look-single profile case. The classifier performance of ETMF,
HMM and hybrid techniques are also shown for 3 and 5 profile averaging. The results
Table I describes the organization of the confusion matrix for 4 target class data set
68
In the Forced decision case, The performance of three different classifiers is compared in
correctly classifying a known target. In table I all the diagonal entries P(BMP2|BMP2),
P(T72|T72), P(2S1|2S1) and P(BRDM2|BRDM2) represents PCC for targets BMP2, T72,
2S1 and BRDM2. More the PCC, the better the classifier is in recognizing the unknown
range profile. Based on table I, the classification performance of Hybrid, ETMF and
Table III: Confusion matrix for ETMF with single profile testing (PCC=81.5%)
69
Target BMP2 T72 2S1 BRDM2
BMP2 0.7731 0.0657 0.0845 0.0767
T72 0.0461 0.9233 0.0213 0.0093
2S1 0.0365 0.0134 0.8698 0.0603
BRDM2 0.0868 0.0378 0.0690 0.8364
Table V: Confusion matrix for Hybrid algorithm with single profile testing (PCC=85.55%)
Table VI: Confusion matrix for ETMF with three profile average testing (PCC=91.035%)
Table IX: Confusion matrix for ETMF with five profile average testing (PCC=94.5675%)
70
Target BMP2 T72 2S1 BRDM2
BMP2 0.8271 0.0334 0.0750 0.0645
T72 0.0398 0.9226 0.0222 0.0154
2S1 0.0305 0.0122 0.9105 0.0468
BRDM2 0.0505 0.0208 0.0955 0.8332
Table X: Confusion matrix for HMM with five profile average testing (PCC=87.3350%)
In tables III-XI the individual performance confusion matrices for ETMF, HMM and
Hybrid classifiers are displayed in single profile, 3 profile averaging and 5 profile
averaging environment respectively. The row targets of the confusion matrices denote the
train templates and the column targets represent observation profiles. In the Forced
Decision case, as it is assumed that each observation sequence belongs to any of the
known target templates, the major disadvantage is it’s inability to distinguish non-targets
from real targets. So Forced Decision classifier has limited practical uses. Though very
little application in airborne radar, Forced Decision is often used to measure strength of
ATR algorithms. From the Forced Decision result, it can be concluded that HMM
classifier performance is not at per compared to ETMF or hybrid classifier. As each HRR
profile contains large variability, longer sequence of profiles is needed to train HMM.
This is quite obvious when we observe table II. HMM classification rate improves from a
meagre 66.67% to high 82.17% (an improvement of about 15.5%) as we proceed from
71
single test profile classification to three profile average classification. Averaging or
sequencing of waveforms improves HMM classifier performance but then also the
120
100
80
HMM
PCC
60 ETMF
ETMF+HMM
40
20
0
1 3 5
No. of Profiles
Fig. 2.19: Bar plot representation of ETMF, HMM and Hybrid classifier performances (In the x-axis, 1,3 and 5
represents classifier performance on single profile, 3 profile average and 5 profile average observation profiles
Figure 2.19 shows a bar plot depicting the classifier performance comparison
between ETMF, HMM and Hybrid techniques. It can be seen that as the observation
72
profile averaging increases performance enhancement of the hybrid algorithm becomes
In this case, the hybrid algorithm is applied to make classification decisions in the
unknown target scenario, which is simulated by rotating the target class using the LOOM
[10] approach. Thresholds for distinguishing between known and unknown targets are
generated according to LOOM approach explained in [11-12]. As the HMM classifier did
not perform as well as ETMF in Forced decision case, it is regarded that HMM classifier
is not at all a major force here and the hybrid classifier performance is compared with
that of ETMF only. The performance results of the classifier are summarized by ROC
Alarm (Pfa). Pd is the probability that a target declaration is made provided a known target
is observed. Known target observations classified as unknown are termed as “Miss” and
denoted as Pmiss.
Pmiss = 1- Pd (2.62)
made. Pfa is the Probability of classifying unknowns as known target. The primary aim of
any ATR algorithm is to minimize Pfa. The Hybrid and ETMF Classifier performances
are also shown through confusion matrices. Table XII describes the organization of the
confusion matrix for 4 target class data set comprising of BMP2, T72, 2S1 and BRDM2.
73
The performance comparison of Hybrid and ETMF ATR classifications is made
through confusion matrix like the one in Table XII. It is distinguishable from the Forced
decision confusion matrix shown in table I in the sense that an extra column and row
labeled Unknown in table XII. The Unknown column represents Miss decisions (Pmiss),
i.e. when the input unknown observation profile was tested against all the templates, none
of resulting discriminant scores were above the LOOM threshold. The Unknown row
represents the targets, which has no signature present in the training dataset. LOOM [22]
is responsible to create the unknown row. The most desirable results would be zeros in
the entire Unknown column and row entries except the lower right-hand corner, which
should be 1 i.e. unknown targets are 100% classified as unknown. This means that all of
the unknown observations were rejected and all of the known observations were
74
From the confusion matrix template shown in table XII, Pd , PCCCond and Pfa is calculated
as follows:
4 4 4
Pd = ∑ PCCi, j + ∑ ∑ Pmis −idi , j (2.63)
i, j=1 i =1 j=1
where i and j are row and column indices respectively. The upper limit of the summation
in equation (41) indicates there are 4 target templates present in the training dataset.
4
∑ PCCi, j
i, j=1
PCC Cond = ∑ (2.64)
( )
4
∑ 1 − Pmissi ,j = 5
i =1
4
Pfa = ∑ Pfai = 5,j (2.65)
j=1
Based on table XII, performances of Hybrid and ETMF classifiers are displayed in tables
XIII-XVIII.
observations:
Table XIII: Confusion matrix (Unknown rejection threshold about 0.6) for ETMF based classifier
75
Classifier Performance Results for Hybrid single look-single HRR observations:
Table XIV: Confusion matrix (Unknown rejection threshold about 0.6) for Hybrid classifier
Classifier Performance Results for ETMF processed single look-3 profile average
observations:
Table XV: Confusion matrix (Unknown rejection threshold about 0.6) for ETMF based classifier
Table XVI: Confusion matrix (Unknown rejection threshold about 0.6) for Hybrid classifier
76
Classifier Performance Results for ETMF processed single look-5 profile average
observations:
Table XVII: Confusion matrix (Unknown rejection threshold about 0.6) for ETMF based classifier
observations:
Table XVIII: Confusion matrix (Unknown rejection threshold about 0.6) for Hybrid classifier
The confusion matrices shown in table XIII-XVIII are generated by fixing the unknown
target rejection threshold at about 0.6 i.e. in 60% of the cases unknown targets are
regarded as unknown.
From each set of confusion matrix, performance parameters PCCCond, Pd and Pfa
are calculated. By varying the unknown rejection threshold over a wide range, a number
of confusion matrices and it’s associate parameters can be generated. The performance
results are expressed in terms of the parameters PCCCond, Pd and Pfa and shown via ROC
77
PCCCond
Fig. 2.20: ROC curves for Probability of declaration vs Conditional Probability of Correct Classification
Fig. 2.21: ROC curves for Probability of False Alarm vs Probability of declaration
78
Since there are three different performance parameters PCCCond, Pd and Pfa, two ROC
curves are drawn to quantize the performances of Hybrid and ETMF. The First one is
showing Pd vs Pfa and the second one is Pd vs PCCCond. In Pd vs Pfa curve, the desirable
curve, the acceptable performance will be to acheive high PCCCond with Pd. Figure 2.20
and 2.21 depicts the classification performance on single profile. From Figure 2.20, it can
be seen that if the observed profile is identified as target, the probability that the Hybrid
algorithm classifies it in the correct target group is more than that of ETMF algorithm.
But in Figure 2.21, the hybrid algorithm generates more False Alarm at a fixed
advantage of the Hybrid Classifier is limited to the known target cases only and moreover
it often makes mistakes in classifying unknown objects. This result is not surprising as
profile. The performance degradation of HMM on single profile is also seen in Forced
Decision case. So single profile classification is actually not the correct way to compare
the classification strength of the Hybrid algorithm. Hence it is decided to plot ROC
curves by forming the test profile using 3 profile averages and 5 profile averages.
79
PCCCond
Fig. 2.22: ROC curves for Probability of declaration vs Conditional Probability of Correct
Classification (each test profile is formed by averaging 3 observed range profiles)
Fig. 2.23: ROC curves for Probability of False Alarm vs Probability of Declaration (each test profile
is formed by averaging 3 observed range profiles)
80
PCCCond
Fig. 2.24: ROC curves for Probability of declaration vs Conditional Probability of Correct Classification
(each test profile is formed by averaging 5 observed range profiles)
Fig. 2.25: ROC curves for Probability of False Alarm vs Probability of Declaration (each test profile is
formed by averaging 5 observed range profiles)
81
PCCCond
Fig. 2.26: ROC curves for Probability of Declaration vs Conditional Probability of Correct
Identification (Single profile testing, average of 3 profile testing and average of 5 profile testing)
Fig. 2.27: ROC curves for Probability of False Alarm vs Probability of declaration (Single profile
testing, average of 3 profile testing and average of 5 profile testing).
82
PCCCond
Fig. 2.28: ROC curves for Probability of False Alarm vs Conditional Probability of Correct
Classification (Single profile testing, average of 3 profile testing and average of 5 profile testing).
Figure 2.22 and 2.23 shows the ROC comparison of Hybrid and ETMF algorithms on 3
profile average observation sequence. Here also the Hybrid technique generates more
False alarm at a fixed probability of declaration compared to that of ETMF. But the
encouraging part is that the false alarm gets reduced compared to what is observed in
Figure 2.21 at same Pd. That is what we expected. As the HMM classification
performance improves with profile averaging or profile sequencing, the hybrid algorithm
should improve it’s false alarm performance over profile averaging. The encouraging
in Figure 2.25, finally the hybrid classifier Pfa vs Pd ROC curve overtakes that of ETMF
classifier. Though the improvement is minor, but the trend is clearly visible. If we go
more and more test profile averaging, the hybrid classifier will outperform ETMF
83
classifier in a large scale. But we limit ourselves upto 5 profile averaging, as one of our
objective was to develop a classifier which will perform joint target tracking with
Next, the performances of Hybrid and ETMF classifiers are summarized in tabular forms.
Table XIX: Conditional Probability of Correct Classification of Hybrid and ETMF classifiers at Pd = 0.7
Table XX: Probability of Declaration of Hybrid and ETMF classifiers at Pfa = 0.4
84
Figure 2.26 and 2.27 summarizes the performance parameters of Hybrid and ETMF on
single profile testing, 3 profile average testing and 5 profile average testing. In Figure
2.28, it is tried to compare the false alarm performance with Conditional Probability of
Correct Classification.
Table XXI: Conditional Probability of Correct Classification of Hybrid and ETMF classifiers at Pfa = 0.4
From table XIX and XXI it can be concluded that, for unknown target scenario at
constant false alarm, the hybrid technique with single profile achieves similar level of
performance as that attained by the ETMF algorithm using at least three profiles. These
results are promising as to develop a system that can perform joint tracking with
recognition, there is not much chance to perform classification on averaged profiles and
even if profile averaging can be done, it has to be done based on limited number of
profiles. So we are searching for an algorithm which can correctly classify targets as well
as that of other algorithms but with less number of profile averaging or sequencing. The
proposed Hybrid algorithm fulfills the criteria. From table XX it can be observed that at
constant false alarm, the probability of detection of the hybrid algorithm is much worse
85
compared to that of ETMF when target decision is made based on single profile. But as
profile averaging increases, the performance of hybrid is getting better and ultimately if
target decision is made on 5 profile average, the hybrid algorithm outperforms the ETMF
as far as probability of detection is concerned. The tables XIX, XX and XXI are
calculated based on a single operating point from the ROC curves shown previously. The
specifications assumed here as Pd = 0.7 and Pfa = 0.4 are arbitrary. Any other
specifications can easily be met by moving around on previously defined ROC curves. If
the specifications cannot be met from those ROC curves, one can consider
the need to perform target recognition in the Aircraft with limited resources. The ETMF
classifier, we estimated is pretty fast and can be made operational in real time. That’s the
disadvantage of the hybrid classifier. As ETMF is combined with HMM and HMM takes
longer time to compute it’s likelihood, Hybrid classifier demands more computing power
The actual computation of the discriminants for the ETMF classifier varies
between 0.0050 s/signature to 0.0160 s/signature. Mostly the computation time in ETMF
86
clustered around 0.0460 s/signature. In summary, we can tell the ETMF-HMM Hybrid
classifier is about 4.2 times more computation intensive than ETMF alone. The reason
behind for such delay of Hybrid classifier is due to HMM Forward recursive likelihood
computation. Hence to implement the proposed Hybrid algorithm in real time, high
2.4.5: Conclusion
The main objective of this research was to demonstrate that improved classification of
demonstrated that the proposed hybrid ETMF-HMM technique improves PCC when
compared to what is achievable by any one of the algorithms applied individually. It was
also demonstrated that instead of classifying with a single profile, the performance of the
making classification decisions. For forced-decision case, our results show that the hybrid
the usefulness of the Hybrid algorithm can be understood from the results obtained in
unknown target scenario. For unknown target scenario, the hybrid technique with single
profile achieves similar level of performance as that attained by ETMF algorithm using at
least three profiles. The results of this research will be integrated into current simulation
efforts for the research and development of a joint tracking based ATR system.
87
3: TIME RECURSIVE MULTIPLE HYPOTHESIS TESTING
It was seen that averaging over several profiles or multilook processing improves the
ATR classifier performance [7]. The multilook approach [7] is quite renowned to
improve the target recognition rate compared to single profile ATR. The main
joint tracking with continuous target recognition for moving targets, the single look
ETMF needs to be applied time recursively as new profiles are observed over time. Here
we propose a time recursive sequential updating technique that will improve the ATR
performance and as well as will facilitate tracking. For sequential processing, when
multiple HRR profiles are observed over time, no particular target decision is taken based
on a single profile. Instead the maximum matched correlations for each target are saved
and combined later with those for other profiles coming in sequence. The main
characteristics for this sequential multiple hypothesis technique is the hypothesis score
for range profiles is always updated in a time recursive manner, hence the target decision
made at time t is dependent on all the hypothesis score from initial upto time t. To
facilitate tracking it is preferable to convert the correlations into probabilities, which can
88
Let c1,c2,…cK be the discriminant scores for the K targets. If the discriminants are
correlation scores, the corresponding probability measures at the t-th time instant can be
found as,
ck
Ptk = K
, k = 1, 2,...K (3.1)
∑c
k =1
k
Over a small time window N, the chosen ATR algorithm determines the target hypothesis
1,…,t-N+1 can then be combined to form the final multiple hypothesis measures for each
target as follows:
t
k
Pt|t,...,t − N +1 = ∏
l = t − N +1
Plk , k = 1, 2,..., K (3.2)
Where, N is the window of the time sequences. This technique is different than profile-
test profiles.
89
Classification based on combined hypothesis score:
The classification ID decision ( It|t,...,t − N +1 ) at the t-th time instant would correspond to the
Here, it can be observed that the classification decision at time t depends not only on the
hypothesis score for the range profile at time t, but also on all the stored hypothesis score
As a new range profile is observed at time (t+1), the hypothesis can be updated as,
For moving targets, it is important to use only a small window of range profiles as
considerably earlier observations may bias the decisions. The multiple hypotheses that is
Ptk+1
Ptk+1|t +1,...,t − N = Pt|t,...,t
k
− N +1 , k = 1, 2,...K (3.5)
Ptk− N +1
The multiple hypothesis probabilities can be utilized by the tracker as confidence level
for each target under track. The block diagram next describes the time recursive multiple
90
1
P t
Single profile ID 2
P
Range Profile at time t
Max
I t |t
t
M 1
1 Maximum P K P k | k ,K , k − N + 1
Templates t
Matched KN discriminants
2
Filter for K targets
discriminants being tracked
M
Hypothesis
P k | k ,K , k − N + 1
Range Profile at
Max M
time t-1
1 I t −1|t −1 M Sequential
K
Maximum Hypothesis Multiple P k | k ,K , k − N + 1
Train Matched Hypothesis
models Filter
discriminants Combiner
KN for K targets 1
being tracked Pt− N +1
Max
Range
profiles
Range
P t2− N + 1 I t |t ,K , t − N +1
Profile at
Max
It−N+1|t−N+1 M
time t-N+1
1
Maximum
P tK− N + 1
Matched
Filter
discriminants
Templates for K targets M
KN
being tracked Hypothesis
The strength of the proposed time recursive multilook approach is tested in two different
ATR classifiers: ETMF based classifier and ETMF-HMM Hybrid classifier. Also the
two-classifier algorithms are allowed to run on two different MSTAR data sets in order to
prove that the proposed time recursive multiple ID updating approach is independent of
91
3.2.1: ETMF classifier Simulation results
The raw HRR profiles used in these studies were formed from SAR imagery provided to
Wright State University by Air Force Research laboratories (AFRL), WPAFB, Dayton,
OH [13]. These Data were collected under the MSTAR program and were distributed in
the form of two separate CDs to be used for template formation (17o) and testing (15o).
The data set contains SAR images of twenty-two military vehicles at two different
depression angle mode 15o and 17o over a total of 360o aspect angle. The target set
includes nine distinct military vehicles namely 2S1, BMP2, BRDM2, BTR60, BTR70,
D7, T72, ZIL131 and ZSU234 in addition to three variants of BMP2 and eleven variants
of T72. These images were converted to a sequence of HRR waveforms through the
IFFT operation in the cross range domain. The HRR signatures have a bandwidth of
The 17o depression angle images from nine of the twenty-two targets were used to
construct the HRR classifier template profiles: 2S1, BMP2 (SN #9563), BRDM2,
BTR60, BTR70, D7, T72 (SN #S7), ZIL131 and ZSU234. All the military vehicles used
The HRR test profiles were generated from MSTAR target images taken at the
15o depression angle. The Eigen Template based ATR classifiers were tested on all
twenty-two targets covering the whole 360o aspect angle. Test targets, which are directly
included in the train database are called direct representative targets. The targets BMP2
(SN #9566), BMP2 (SN #C21), T72 (SN #132), T72 (SN #812), T72 (SN #A04), T72
92
(SN #A05), T72 (SN #A07), T72 (SN #A10), T72 (SN #A32), T72 (SN #A62), T72 (SN
#A63) and T72 (SN #A64) were not included in the training database but have
representative signature in the training set. These targets are referred as independent
targets.
The ETMF classifier performance is evaluated among the single profile hypothesis
(N=1), 3 profile hypothesis (N=3) and 5 profile hypothesis (N=5) through a set of ROC
curves shown in Figure 3.2-3.3. The time window size N is kept small as considerably
PCCcond
Fig. 3.2: ROC curves for Probability of detection vs Conditional Probability of Correct Classification
93
Fig. 3.3: ROC curves for Probability of False Alarm vs Probability of declaration
Figure 3.2 provides the relationship between Probability of declaration (Pd) and
conditional PCC. It is observed that for a fixed probability of declaration, (PCC)cond for
multiple profile hypothesis ETMF is greater than that of single look ETMF. Figure 3.3
a test profile as target and probability of false alarm. The primary objective of all
automatic target detection schemes is to minimize the false alarm while improving the
probability of correct identification. It is observed here, for a fixed false alarm Pfa=0.4,
(Pd)single look = 60.52%, whereas (Pd)3 profile recursive = 76.85% and (Pd)5 profile recursive =78.59%.
Considering the performance of all the ROC curves it can be concluded that the proposed
94
time recursive multiple profile hypotheses improves the classifier performance compared
The Hybrid ATR classifier performance is evaluated among the single profile hypothesis
Hybrid (N=1), 3 profile hypothesis Hybrid (N=3) and 5 profile hypothesis Hybrid (N=5)
through a set of ROC curves shown in Figure 3.4-3.6. The time window size N is kept
Fig. 3.4: ROC curves for Probability of False Alarm vs Probability of declaration (single profile test,
Time recursive 3 profile updating test and time recursive 5 profile updating test)
95
PCCcond
Fig. 3.5: ROC curves for Probability of Declaration vs Conditional Probability of Correct Identification
(single profile test, Time recursive 3 profile updating test and time recursive 5 profile updating test)
PCCcond
Fig. 3.6: ROC curves for Probability of False Alarm vs Conditional Probability of Correct Identification
(single profile test, Time recursive 3 profile updating test and time recursive 5 profile updating test)
96
Figure 3.4-3.5 provides the relationship between Pfa- Pd and Pd-PCCcond. It is observed
that for a fixed probability of declaration, (PCC)cond for multiple profile hypothesis
ETMF is always greater than that of single look ETMF. Figure 3.6 provides a relationship
between Pfa- PCCcond. The primary objective of all automatic target detection schemes is
to minimize the false alarm while improving the probability of correct identification.
time recursive Hybrid techniques (N=3 and N=5) are displayed in the following tabular
form:
Table XXII: Improvement of Probability of Declaration of Hybrid classifiers due to time recursive
multilook approaches at Pfa = 0.4
97
Considering the performance of all the ROC curves it can be concluded that the proposed
time recursive multiple profile hypotheses improves the classifier performance compared
to single look ATR technique. The time recursive profile updating technique is of great
help to tracking as here target ID is updating recursively with time and no profile
averaging is required.
3.3: Conclusion
It was demonstrated that instead of classifying with a single look, the performance of the
important for performance improvement if the target is moving and changes position
instantly with time. As profile averaging is difficult for moving target, the only way to
proposed in this thesis. The results of this research will be integrated into current
simulation efforts for the research and development of a joint tracking based ATR
system.
98
4: IMPROVED SAR TARGET DETECTION USING SUBSPACE
FILTERING
employs physical and statistical models is developed and presented. This joint
physics/statistics based technique generates realistic images that have many of the “blob-
like” and “spiky” clutter characteristics of UWB radar data in forested regions while
Firstly, the developed SAR target detection algorithms are trained in offline mode
to learn the clutter statistics. But as it is not always possible to learn beforehand,
approaches towards developing “self-training” algorithms for UWB radar target detection
are also investigated using the results of this simulation process. These adaptive
algorithm where small blocks of data in the neighborhood of a sliding test window are
99
4.1: Introduction
Synthetic Aperture Radar (SAR) imagery is commonly used for detecting, classifying and
important yet challenging problem especially if the target is hindered under outliers. In
previous work, it has been proposed a number of approaches to target detection [64-65]
in the ultra-wideband (UWB) SAR area. The approaches presented include detailed
proposed in [64-65]. An approach to developing a new class of rank order filters, known
as, “discontinuity filter” for ultra-wideband radar target detection applications was
algorithms that are designed for basic UWB SAR sensor phenomenology but do not
require extensive off-line training. Both of these approaches have been shown to
generate acceptable levels of performance under certain conditions that are of interest for
This section of the thesis presents theoretical development and its associated
simulation results of a new ultra-wideband SAR target detection algorithm in offline and
procedures, where a set of localized regions within a given SAR image are sampled in
real-time for purposes of obtaining low-order and robust real-time clutter statistics. The
real time clutter models are applied in a sliding-window type target detection paradigm
for clutter cancellation and target detection. Results are presented from the analysis of
100
three new categories of algorithms that were developed for this investigation, denoted as
generate satisfactory results for severe UWB SAR impulsive-type clutter. Initially, all
three algorithms were allowed to run in offline training mode and later concentration is
characterized as, no offline training is required and the algorithm will learn as it flies on
adaptive and self-trained target detection algorithms are reported. The results indicate that
eliminate severe clutter discretes that adversely effect performance, appears promising
algorithms are given emphasis because in some applications the target detection classifier
needs to learn while “on the fly” due to a lack of accurate off-line training data.
Simulated SAR impulse clutter is generated by modulating a “top hat” model for the SAR
video phase history with a K-distributed amplitude. Targets are synthesized within the
Figure 1 shows a block diagram of the UWB SAR simulation developed for this
investigation. Two types of clutter were embedded into the SAR image: forest clutters
and residual background clutters. The central scattering structure for UWB forest clutter
101
elements are modeled by applying a physics-based parametric radar scatter center model
102
The general scattering model is described by [66]
AHH
k
α
n
HV f i 2πf 4πf
E(f , φ) = ∑
k =1
Ak j sin c
AVV fc c
Lk sin ( φ − φk ) exp − j
c
[ xk cos φ + yk sin φ] (4.1)
k
Here we are assuming that the wavelength of the incident excitation is much smaller
compared to the target extent, so that the effective backscattered field is the vector sum of
the entire backscattered field emitted from each scatter center. The backscattered field E
In equation (1), E (f, φ) is the set of backscattered field in all 3 polarization’s HH,
{A k , x k , y k , α k }kn =1 characterize the n individual scattering centers. For the kth scattering
center, (xk,yk) gives the slant plane location with respect to a zero-phase reference, αi
characterizes the geometry and Ak is a complex scalar characterizing the magnitude and
phase. The round trip propagation delay from each kth scattering center is denoted as:
For this application, targets are modeled as long dihedrals and clutter scattering
centers are modeled as short tophats. For a single-dihedral target model with HH
polarization, A HV
k
and A VV
k
are zero, and αi is equal to one . The synthetic images were
created using Ohio State’s models, followed by SAR image formation. Tree trunks were
modeled as tophat reflectors and targets were modeled as varying length dihedrals. The
forest was created by randomly placing trees in the image. The reflectivity of the trees
are K-distributed, and the locations are uniformly distributed. The tree canopy was
modeled with a Gaussian distribution with the mean and the canopy radius both
103
proportional to the trunk amplitude. A 2D Hanning window was used to smooth the
transition between the canopy and the background. Finally, Gaussian noise was added to
dependence in the argument of the sinc term in equation (4.1) is removed and the sinc is
varied as a function of the dihedral length, L, only. The total number of synthetic
scatterers in the SAR video phase history (VPH) computation is denoted by n. As shown
in the VPH block of Figure 4.1, the VPH is synthesized by breaking the summation in
equation (4.1) into two parts and then selecting the parameters for n scatterers: long
dihedrals for targets (15 targets is the default number with length varying from 0.5 to 1.5)
and tophats for clutter (n-15 number clutter scatterers). Here the chosen n is 600, a
default number. The emphasis was given to have a high clutter to target scatterer ratio,
which is 20:1 in our simulated image. The x, y locations of the clutter scatterer centers
In addition, for this simulation approach, we make the heuristic argument that the
scattering, etc.) is random. This allows us to generalize statistical arguments from the
statistical signal processing literature [67] that indicate that, if the number of scatterers in
an “elementary cell” is finite and random, the number of scatterers is roughly Poisson
distributed. The expected value of this Poisson distribution is typically also assumed
random with a gamma distribution, and the resulting intensity distribution is typically
assumed K-distributed. K-distribution was chosen for clutter intensity modeling due to
the assumption that neighboring clutters are correlated. For this simulation, we generalize
104
the concept of an “elementary cell” to include the region in the immediate local
due to the neighboring canopy region. The parameters of the K-distributed intensity are
randomly distributed to generate a severe clutter environment that has many instances of
impulsive “spikes” and “blobs”. To this end, a challenging simulation environment was
electromagnetic computations.
collection is not needed. The software-based capability allows us to vary clutter severity,
the number of targets and separations between targets etc. to resemble any desired
realistic UWB images. Hence, it is possible to test the effectiveness of various target
detection algorithms for many variations of target and clutter operating scenarios.
The x, y coordinates of the VPH model in equation (4.1) are further synthesized
and transformed into airborne SAR range coordinates and then processed by a wideband
After SAR image formation with wavefront processing, the resulting SAR image is post-
processing for additive noise and for local variations in the residual clutter level in
The radar signal attenuation due to foliage can be summarized for the case of typical
sin π 4
Two-Way Attenuation (dB) = β f α (4.3)
sin γ
105
where, f is the frequency in Megahertz, γ is grazing angle, and, for HH polarization, the
median attenuation is for α = 0.79 and β = 0.044 and the 90th percentile attenuation is for
α = 0.79 and β = 0.077. The α and β values are chosen according to the forest clutter
Two samples of UWB SAR imagery generated with this simulation process are
illustrated in sub-section 4.5 for the purposes of testing and training. In the next two sub-
sections the SVD operation on SAR images and its effectiveness in SAR clutter
Given an N x N SAR image data window Y, basic singular value decomposition (SVD)
M
SVD
Y → UΛV T = ∑λ u v
i =1
i i
T
i (4.4)
where,
Singular Values :
The EV[·] operator represents “the eigenvectors of” the matrix inside the brackets. For
Range Vs. Crossrange SAR data, the left eigen vectors (U) span the orthogonal basis
106
space in the range domain while the right eigen vectors (V) span the Crossrange domain.
λ11 ≥ λ 22 ... ≥ λ NN , where λ ii denotes the weights associated with i-th eigenvector.
A critical problem faced in ATR is clutter noise that often corrupts SAR images.
sonar, speech, image etc., where SVD has been found to be highly effective for noise
suppression and also for improved estimation of underlying system parameters. The
the “Signal subspace”, whereas those corresponding to the smaller singular values are
common approach has been to reconstruct the original matrix, after "zeroing-out" what
are considered to be the noise or clutter subspaces i.e. the clutter reduced HRR matrix can
be reconstructed as,
N
ˆ =
Y ∑ λˆ uˆ vˆ
i =1
i i i
T
(4.8)
where, λ̂ i is the estimated set of dominant eigen values, ûi and v̂ i are the estimated set
Filtering. In the next section we discuss the three SVD-based algorithms that are
developed to adaptively cancel UWB radar impulse clutter based on the approach of
107
4.5: SVD based SAR Target Detection Algorithms
Theoretical approach of the three proposed SVD based algorithm EN-SND, CS-SVD and
4.5.1.1: Motivation
The proposed algorithm for energy normalized SVD technique is based on Eigen
filtering. Eigen filtering is an effective way to decouple target from clutter. In this
algorithm the sum of the ratio of eigenvalues between localized windows are taken as a
metric to discriminate between target and clutter. This idea is taken from the
homogeneity test, described in [70] where it is shown that the generalized inner product
(GIP), which is a test of eigen structure and detecting forms of non-homogeneity is the
sum of the ratio of the eigenvalues of the range-crossrange space correlation matrix.
Homogeneity test was chosen as the discrimination factor between target and clutter due
to the homogeneous nature of clutter scatterers i.e. energy of the clutter scatterers are
uniform, whereas targets are mostly rank one i.e. the highest eigenvalue contains
significant amount of energy compared to those lesser ones. In UWB environment, the
came from the plotting of eigenvalues of a 15 × 15 target only image block with that of
same size clutter block. From Figure 4.2, it can be verified that in case of target, the
highest eigenvalue contains significant amount of energy compared to it’s smaller ones,
but in case of clutter, atleast 4-5 eigenvalues are necessary to completely describe the
108
clutter blob. Hence for this reason, the magnitude of the highest eigenvalue of the target
blob is far too high compared to that of clutter blob. This information can be used as the
discriminant factor between target and clutter image block and the idea of the proposed
eigenvalue ratios from pairs of neighboring image windows. Statistics from this
“training” histogram are used to determine a threshold. During the “testing” phase, sums
of eigenvalue ratios are computed for each localized test clutter window with respect to
the localized test window with the minimum energy (i.e. minimum sum of eigenvalues).
assumed to be lower than that of target. If the test “sum of eigenvalue ratio” metric is
below the train threshold, the value in the given test window are “zeroed out. If this test
metric exceeds the train threshold, the data in the test window is reconstructed from the
pre-dominant SVD eigenvalues and eigenvectors. This process attempts to eliminate the
109
Fig. 4.2: Eigen Spectrum of Target and Clutter blob
The detailed steps for the training phase are outlined as follows:
1) The train clutter only image is processed as segmented 3m × 3m regions that are each
Figure 4.3 shows the image segmentation and feature extraction process in graphical
manner. A11, A12..., A33 are the respective feature sets extracted from each m × m region
110
A11 A12 A13
m
A22
m
A21 A23
3m
A31 A33
A32
3m
2) The highest m/3 eigenvalues are calculated for each m × m window within a given
m/3 eigenvalues contains more than 95% of energy irrespective of target or clutter in
each m × m region. In the EN-SVD approach, as sum of ratio of the eigen values are
chosen to be the discriminant factor between target and clutter, hence eigen values are
treated as feature components and the feature vectors A11, A12..., A33 each contains
111
3) Sum of the ratio of A11 with all other (n-1) feature vectors are calculated. This yields
A A A A
B1 = ∑ 11 , ∑ 11 ,..., ∑ 11 ,..., ∑ 11 (4.9)
A12 A13 A 22 A 33
4) Similarly sum of the ratio of A12 with all other (n-2) feature vectors (n-2, because
A11
is already done in step (3)) are calculated and stored in B2.
A12
A A A A
B 2 = ∑ 12 , ∑ 13 ,..., ∑ 12 ,..., ∑ 12 (4.10)
A13 A 21 A 22 A33
A 32
This process will go on, until we get, B n −1 = (4.11)
A 33
5) All individual vectors B1, B2,..., Bn-1 are concatenated to generate a large B vector.
6) Next, in a similar procedure B vector is created for a number of different clutter only
7) With all the elements of B a histogram is formed and T=mean+sigma (sigma signifies
one standard deviation of the histogram) of the histogrammed output will be denoted
more number of m × m clutter only regions are used to train the EN-SVD classifier,
more robust the classifier will be to discriminate between target and clutter.
112
4.5.1.3: EN-SVD Testing Phase
The detailed steps for the testing phase are outlined as follows:
1) The test SAR image is processed as segmented 3m × 3m regions that are each divided
into n=9, m × m windows same as train-clutter image. The respective highest m/3
eigenvalue feature sets are calculated from each small m × m window and denoted as
energy is found out. Let the feature vector of the least energy window is denoted as
Akl (1 ≤ k ≤ n / 3,1 ≤ l ≤ n / 3) .
3) Sum of the ratio of Akl with all other (n-1) feature vectors are calculated. This yields a
A A A A
C = ∑ kl , ∑ kl ,..., ∑ kl ,..., ∑ kl (4.13)
A11 A12 A 22 A 33
4) Each (n-1) metric value in vector C is compared with the predetermined threshold T
If T ≥ C( j) , 1 ≤ j ≤ n − 1
responsible for generation of C(j) with the lowest energy window. The corresponding
113
Else T ≤ C( j) , 1 ≤ j ≤ n − 1
assumed that the window consists of a whole/part of a target and the data in that
5) The entire operation described in steps (1)-(4) is repeated throughout the whole image
In this approach, the training data is statistically pre-filtered before running the EN-SVD
from section (4.5.1). The need for the CS-SVD can be observed in the simulation section,
due to significant amount of target information loss in EN-SVD case. The problematic
area detected is the presence of large bright clutter blobs in the image, so before
implementing EN-SVD on the SAR test image, the bright clutter blobs should be
Although many possibilities for a statistical filter can be investigated, for this
particular investigation a simple mean filter is applied. The main intention here is to
observed that mxm window is sufficient to capture the whole clutter blob, mean is
computed over a sliding mxm training regions of interest. Then, the mean of the
resulting mean-filter output over all training regions is used as a “blob” threshold for the
training. In other words, any mxm training region whose mean exceeds the mean
114
threshold is denoted as severe “blob-type” impulse clutter and are zeroed out from the
training set. Hence what we get is a filtered train image. Next the training algorithm of
For testing, the same mean prefiltering steps are followed on the observation/test
image before applying the same EN-SVD testing algorithm. As shown in the simulation
section, this approach can allow EN-SVD to more effectively cancel the “typical clutter”
while preserving more targets and can yield enhanced target detection performance.
normalized SVD approach. The motivation of this algorithm remains the same as that of
used to suppress the bright clutter blobs. In this approach a convolutional prefiltering step
distance metric from the kernel center in an effort to pre-filter severe “blob-like” and
“sparsely-impulsive” clutter discretes. The particular filter kernel designed for this
sliding window over the observation/test image. This kernel function is calculated by
setting all pixels within an 11-pixel radius of the kernel center equal to zero and by
equating the remainder of the pixels equal to the Euclidean distance between the kernel
115
center and the pixel. The test image is convoluted over pixel by pixel sliding terrain and
all the convoluted outputs are used to generate a histogram of convoluted outputs.
The primary motivation for the formation of such type of Euclidean mask came
due to the discrete nature of clutter blob. The scattering centers of the clutter blobs are
distant apart unlike that of targets where they are quite coherent. The designated mask,
assuming the target is at center, tried to estimate the density around the target. If target
remains at the center of the sliding mask and there is no bright scatterers around the
target, the convoluted output will be zero as the center target is already masked by the
kernel. Hence if the convoluted output generates a large value, that means there are
discrete scatterers present around the target at large distance (because here we are
multiplying the scatterer intensity with distance from the center) and it is needed to
suppress those discrete blobs in order to improve the target detection performance of EN-
SVD.
116
Figure 4.5 shows the histogram of the data at the output of the terrain prefilter, along with
the “center”, Sigma-1, and Sigma-2 values for this simulation. The resulting histogram
filter data from the test image is processed in accordance with the following procedure:
Sigma1 Sigma2
Center
The maximum value of the filter output histogram distribution is denoted as the “Center”
and Sigma-1/Sigma-2 are standard deviation of all the filtered values to the left/right of
the histogram “Center”. A given test image pixel is set equal to zero if the corresponding
filter output is beyond Sigma-1 to the left of the histogram “Center” or is beyond “Sigma-
2” to the right of the “Center”. It is assumed that those test image pixels correspond to
either blob type clutter or discrete residual clutter. But those test image pixels whose
corresponding filter output is in between sigma-1 and sigma-2 of the histogrammed plot
are kept intact assuming target information with clutter are embedded in those region and
117
EN-SVD will be performed on this image to nullify further clutter and improve target
detection performance.
This filter has the effect of filtering-out a significant number of severe “blob-like”
The overall TF-SVD algorithm can me implemented in the following three steps
described next:
a) The sliding-window terrain filter kernel is applied on the train clutter only image and
b) Center test image pixels corresponding to the image-mask region is set equal to zero
in accordance with histogram Sigma-1 and Sigma-2 criteria as described in the kernel
formation section. We can assume mostly all the clutter blobs and discretes are
nullified after this approach. Then EN-SVD algorithm is applied on the train clutter
algorithm is applied on the same image to eliminate the clutter blobs that might be
c) For testing, the same Euclidean masking steps are followed on the observation/test
image before applying the same EN-SVD testing algorithm. As shown in the
simulation section, this approach can also allow EN-SVD to more effectively cancel
118
the “typical clutter” while preserving more targets and can yield enhanced target
detection performance.
In the simulation section it is also shown that these approaches has potential to act as self
training algorithms i.e. where no separate train image is required to learn the algorithm,
Fig. 4.6: Sample UWB SAR Simulation test Image (target with clutter, where targets are boxed
with red rectangle)
119
Fig. 4.7: Sample UWB SAR Simulation Clutter only image
In this section, a set of comparative results are presented by running the EN-SVD, the
CS-SVD, and TF-SVD algorithms on the sample simulation test image of Figure 4.6. The
synthetic image was created using Ohio State’s models, followed by SAR image
formation. This sample image shows many instances of “spiky” and “blob-like” realistic
UWB radar clutter. The targets, modeled as dihedrals, are shown along three rows on the
top, middle, and bottom of the simulation. There are altogether fifteen targets in the
image, whereas five targets are placed on each row. The bottom row has relatively long
targets (length L=1.5), the middle row has targets of relatively medium length (length
120
L=1), and the top row has short targets (length L=0.5). The target backscatter level or
reflectivity varies from left to right. The reflectivity was varied based on the length of the
targets. The reflectivity was calculated as: f = αL , where α varies from 2 to 10. The
UWB radar clutter was modeled in two phases: 1) bright or blob clutter modeled as tree
trunks and tree canopies 2) residual clutter. Tree trunks were modeled as tophat reflectors
and targets were modeled as varying length dihedrals. The forest was created by
randomly placing trees in the image. The reflectivity of the trees are K-distributed, and
the locations are uniformly distributed. The tree canopy was modeled with a Gaussian
distribution with the mean and the canopy radius both proportional to the trunk
amplitude. A 2D Hanning window was used to smooth the transition between the canopy
and the background. Finally, Gaussian noise was added to represent the noise floor.
All the proposed SVD based target detection algorithm is trained first by feeding
the clutter only image shown in Figure 4.7. From that clutter only image, the SAR target
detection algorithms would learn the clutter statistics and determine the threshold to
distinguish between target and clutter. Then the performance of all three-detection
algorithms is tested in Figure 4.6 and their performance comparison is made through a set
of ROC curves. It is to be noted that the clutter only simulation UWB image (Figure 4.7)
and clutter with target UWB simulation image (Figure 4.6) were generated using two
different simulation runs and conditions are put such that random numbers which
determine position and intensity of clutter blobs/residual clutter varies with different
simulation runs. This was done intentionally to make those two images uncorrelated, as
finally the proposed algorithm will be tested in self-training mode. Each segmented
window size is taken m=15 as the maximum size of a clutter blob is 15x15.
121
4.5.4.2: Performance Comparison of Target Detection algorithms
All three SAR target detection algorithms EN-SVD, CS-SVD and TF-SVD are first
applied on the train clutter only image (Figure 4.7) to generate the threshold, which is the
distinguishing factor between target and clutter. Then all those three algorithms are tested
in the target-clutter image (Figure 4.6) and nullify clutter based on the predetermined
threshold.
Fig. 4.8: UWB SAR simulation image after performing EN-SVD on Figure 4.6
122
The performance of the EN-SVD algorithm on Figure 4.6 is displayed in Figure 4.8.
From the figure, it can be seen that although the algorithm nullifies significant amount of
clutter from the neighborhood of targets, severe target information is also lost. Especially
the targets which have low backscatter energy or have smaller length suffers the most.
The colored boxes are drawn around the remains of the visible target. Those targets,
which can be visible almost completely, are surrounded by a red box and the partial
visible targets are surrounded by yellow rectangles. From Figure 4.8, it can be seen that
only two targets are fully recovered and five other targets are partially recovered after
implementing EN-SVD on Figure 4.6. This is highly unwanted, as the EN-SVD detection
algorithm looses a number of targets almost completely. This will definitely reduce the
performance of EN-SVD and this is the motivation for go beyond EN-SVD and the
algorithms of CS-SVD and TF-SVD come into picture. The reason for worse
performance of EN-SVD is quantified due to the presence of bright clutter blobs whose
reflectivity is comparable with that of targets. Hence because of the high reflectivity
clutter blobs, the threshold obtained from the training process is relatively high, which is
nullifying not only significant amount of clutter but also large portions of targets. Hence
energy clutter blobs are suppressed before applying EN-SVD. The CS-SVD does just
that, it just has a pre-filtering overhead with EN-SVD. Figure 4.9 displays the image after
applying CS-SVD algorithm. From there it is observed that not only it zeros out the
clutter blobs and clutter discretes, but also it preserves significant amount of target blobs
123
Estimated UWB Image
Figure 4.9: UWB SAR simulation image after performing CS-SVD on Figure 4.6
Figure 4.9 is the displayed SAR image after implementing CS-SVD on Figure 4.6. Here
also the partial/full remain targets are boxed with color rectangles. The red rectangle
surrounds those targets whose maximum portions are still visible in the image and the
yellow rectangle covers the targets whose significant portions are removed but still can
be distinguishable as a target. From Figure 4.9, it can be observed that the CS-SVD
124
algorithm retains at least six targets in full and three targets in partial i.e. much
Fig. 4.10: UWB SAR simulation image after performing Euclidean masking operation in TF-SVD
Figure 4.10 shows the image after implementing the TF-SVD pre-masking operation on
Figure 4.6. It is observed that the Euclidean masking operation preserved almost all
targets except the leftmost in the top row. The main importance part here is all the large
bright blob type clutters are nullified from the image. Now the EN-SVD algorithm will
125
be applied on Figure 4.10. As the performance detoriating blob type clutter are not there
Fig. 4.11: Final UWB SAR simulation image after performing TF-SVD
From Figure 4.11, it can be observed that finally the TF-SVD algorithm retains almost all
the targets either full or partial. The targets whose most parts are visible are surrounded
126
by red rectangle and those targets which looses a significant part but still can be
qualitatively. From the qualitative result, it is expected that both CS-SVD and TF-SVD
Fig. 4.12: Performance comparison of EN-SVD, CS-SVD and TF-SVD in offline training-real time
testing mode
127
Fig. 4.13: Performance comparison of EN-SVD, CS-SVD and TF-SVD shown in logarithmic scale in
offline training-real time testing mode
The ROC curves shown in probability of detection as a function of varying the false
alarm threshold in Figure 4.12 and 4.13. In Figure 4.13, the x-axis is plotted in
logarithmic scale. The ROC result confirms the visual inference that CS-SVD and TF-
SVD outperforms the baseline EN-SVD in terms of target detection performance. For
0.66, whereas at the same false alarm probability, CS-SVD provides 0.72 and TF-SVD
128
provides 0.8 as probability of detection. In case of EN-SVD, probability of detection goes
to 1 (100% detection) at a high false alarm rate of 0.78, but in case of CS-SVD and TF-
SVD, 100% detection is achieved only at a small false alarm of 0.2 and 0.22. From the
above ROC plots, it can be concluded that CS-SVD and TF-SVD clearly outperforms the
baseline EN-SVD.
Next, the three proposed SAR target detection algorithm EN-SVD, CS-SVD and TF-
SVD are compared in self-training mode. In the self-training mode, the three algorithms
will make use of only the simulated target-clutter image (shown in Figure 4.6), as it will
learn while on the fly. Here, the assumption is there is no absolute requirement of training
image, as the algorithms will learn on the fly while performing target detection. The great
advantage of the self train algorithm compared to other offline training techniques is 1)
here the algorithm need not be familiar with the environment 2) the algorithm will learn
the clutter statistics while simultaneously performing target detection, hence it performs
efficient time and resource management. Clearly, development of self training algorithm
representative data a priori and the target detection algorithm needs to function in an
unknown environment.
and TF-SVD are analyzed in self-training mode via a set of ROC curves.
129
Fig. 4.14: Performance comparison of EN-SVD, CS-SVD and TF-SVD in self-train mode
From Figure 4.14 it can be seen that the target detection algorithms are compared in local
self train mode and global self train mode. In global self train mode, first the whole test
image is generated. Then the threshold to discriminate target with clutter is computed
based on the whole test image. Lastly the algorithm again works on the image and
nullifies clutter while keeps target. It is to be noted that, as each algorithm works on
segmented images, the threshold remains constant for all image segments here. This is a
disadvantage as the global threshold is based on the clutter statistics of the whole region,
130
it may not reflect the statistics of area-localized clutter. So the threshold obtained may not
be robust for all parts of the image. The primary disadvantage is in this way of threshold
computation the algorithm needs to be worked two times on the image: First time it will
calculate the threshold and then again the control point will be shifted to the starting point
of the image and it starts nullify clutter based on the threshold. Hence this process can
hardly be made as real time. As SAR generation is itself a high time consuming process,
the objective of any target detection scheme based on SAR is to reduce further overhead.
To combat with this requirement, the scheme to generate threshold locally while
performing target detection is proposed. That means, one threshold is computed for every
segmented portion of the image. Each algorithm first computes the threshold from a small
mxm segmented portion and then removes clutter from that segmented portion based on
the threshold calculated before. The main advantage is each localized area has it’s own
threshold i.e. the threshold is determined based on localized clutter statistics, hence it
should be more effective compared to global threshold. Next, as each algorithm just
needs to scan the whole image once, a lot of computation time is saved and makes this
eagerness was to see their performance in self train mode. Figure 4.14 displays that. As
expected, it can be seen, the local self train mode outperforms the global self train
algorithms. The performance of EN-SVD in local self train mode does not cost any extra
gain over global self-train mode and same as the case for CS-SVD, where their
performance are almost comparable. But the noticable improvement observed, if we see
the performance curves of TF-SVD in local self train mode comapred to that in global
131
self train mode. At probability of false alarm of 0.05, the probability of detection of TF-
SVD in global self train mode and local self train mode is 0.8 and 0.88 respectively. The
TF-SVD algorithm in local self training mode provides 100% target detection at a small
about 80.5%. Hence we propose to implement the TF-SVD with local threshold in SAR
Fig. 4.15: Performance comparison of EN-SVD, CS-SVD and TF-SVD in self-train mode (here x –
axis in logarithmic scale)
132
4.5.4.3: Performance Comparison of various techniques
133
Fig. 4.17: ROC Performance comparison of various techniques (x axis is in logarithmic scale)
offline train mode and self train mode together in a single picture. Figure 4.16 displays
that. In addition to that there are median filtering techniques combined with CS-SVD and
TF-SVD cases i.e. CS-SVD and TF-SVD output images are meadian filtered before
scoring the targets and clutter to form ROC curves. Also only the median filter output is
134
also shown i.e. the test image is median filtered and then target detection algorithm ran
onto it to generate ROC curves. In addition, a different image segmented size m=5 is
chosen in CS-SVD offline case to verify that the performance of CS-SVD algorithm will
be optimum at m=15. It is to be noted that median filtering is not implemented with none
of the algorithms that are running in self training mode. This is intentionally done to keep
the overhead minimum as the self training mode runs in real time and median filtering
slows the system. Another reason for this is median filtering hybrid with one of the
compared to that algorithm alone. This can be verified if we observe the target detection
performance in CS-SVD offline case with that of CS-SVD + median offline case. Same is
true for TF-SVD also. It can be verified from the above plots that the proposed CS-SVD
and TF-SVD significantly improve the target detection performance at low false alarm
rate compared to median filtering used for target detection from SAR image.
4.5.4.4: Conclusion
In summary, this initial research in the area of “self-training” algorithms for UWB SAR
target detection in a severe clutter simulation environment indicates that there is potential
pre-filtering stages that eliminate significant portions of hard clutter discretes associated
with many severe UWB radar environments. The simulation ROC curves and images
before applying the baseline algorithm onto it. The Terrain Filtered SVD approach
135
train mode. Possible scope for further research includes developing approaches that, for
136
5: SUMMARY AND FUTURE WORK
5.1: Summary
A new approach using High Range Resolution (HRR) profiles to recognize time critical
military targets is considered in the first part of the thesis. In the second part, a new time
with recognition, is proposed. In the last part, three new SAR target detection algorithms
ATR in recognition of moving targets, the HRR scheme, studied as part of this thesis, can
computational savings when compared with SAR-ATR scheme. Results described in this
This thesis includes details of how HRR profiles are created and pre-processed to
achieve superior ATR performance. For large number of target classes with significantly
137
Many commonly known estimation and detection algorithms for ATR work best
or are computationally simple for gaussian signals. HRR profiles being positive valued,
their density function tend to be Gamma or Rayleigh distributed. The power transform
(PT) operation is shown to transform the density function to more Gaussian that in turn is
shown to result in higher recognition rates for uncorrupted Observation data. But it was
HRR data leads to considerable deterioration in the ATR performance. Hence in this
thesis, the power transform operation is avoided while making HRR profiles
was shown to be a pre-requisite for the use of Matched filter in the classification stage, to
In this thesis, a new ATR algorithm is proposed, which will combine the Eigen
Template based Matched Filtering (ETMF) technique with Hidden Markov Model
(HMM) and target classification is performed based on this hybrid technique. The
described in detail. The effectiveness of SVD for removal of clutter from a HRR profiles
are described next. It has been shown that Eigen filtering technique, i.e., reconstruction
after zeroing out small singular values, which correspond to noise or clutter, can be
highly effective removing clutter. Next, the target classification using HMM technique is
described in detail. Lastly, the theoretical approach is provided for the process of
combining between ETMF and HMM. The experimental results are provided in Forced
138
5.1.2: Time Recursive Sequential ATR
In the time recursive sequential ATR technique, target decision is not based on a single
profile. Instead the maximum matched correlations for each target are saved and
combined later with those for other profiles coming in sequence. The target decision is
based on the combined hypothesis score. The main characteristics for this sequential
multiple hypothesis technique is the hypothesis score for range profiles is always updated
in a time recursive manner, hence the target decision made at time t is dependent on all
the hypothesis score from initial upto time t. The theoretical approach to develop this
algorithm is presented along with simulation results comparing single profile hypothesis.
that employs physical and statistical models is developed and presented. The three
proposed SAR target detection algorithms are named as EN-SVD, CS-SVD and TF-SVD.
Firstly, the developed SAR target detection algorithms are trained in offline mode to
learn the clutter statistics. But as it is not always possible to learn beforehand, approaches
towards developing “self-training” algorithms for UWB radar target detection are also
investigated using the results of this simulation process. These adaptive approaches
employ some form of modified Singular Value Decomposition (SVD) algorithm where
small blocks of data in the neighborhood of a sliding test window are processed in real-
time to estimate clutter characteristics. These real-time clutter models are then used to
cancel clutter in the sliding test window. In this thesis, the theoretical approach to
139
develop the algorithm is described first and then the algorithm is analyzed with
simulation results.
The algorithms proposed in this thesis are made with the objective to achieve high
ATR/Detection for moving targets using minimum processing time. Recent advances in
superior sensor technologies and sensor simulation tools that allow wider classes of target
scenarios available at the ATR developmental stage, higher resolution imaging based on
appropriate advanced ATR strategies are all expected to be beneficial for achieving
processing time. Due to the transform it is possible that the energy of some targets may
be distributed among more than the largest singular value and more than one eigenvector
may be more appropriate for template formation for some targets to achieve reasonable
ATR results.
In target detection case, further research includes developing approaches that, for
140
BIBILIOGRAPHY
[1] Li, H.J., and Yang, S.H. (1993), “Using range profiles as feature vectors to identify
aerospace objects”, IEEE Transactions on Antennas and propagation, 41, 3 (1993), 261-
268.
[2] Hudson, S., and Psaltis, D. (1993), “Correlation Filters for aircraft identification from
radar range profiles”, IEEE Transactions on Aerospace and Electronic Systems, 29, 3
(1993), 741-748.
[3] Mitchell R.A. and Westerkamp J.J. (1999), “Robust Statistical feature based aircraft
1077-1094.
[4] Chiang, H-C., Moses, R.L. and Potter, L.C. (2000), “Model based classification of
[5] Novak, L.M., Owirka, G.J. and Weaver, A.L (1999), “Automatic Target Recognition
using enhanced resolution SAR data”, IEEE Transactions on Aerospace and Electronic
141
[6] Mirkin M. and Hodges, B. (1997), Private conversation at MIT Lincoln Laboratory,
1997 (according to Nguyen D., Benitz G., Kay J., Orchard B. and Whiting R.).
[7] Nguyen, D., Benitz, G., Kay, J., Orchard, B. and Whiting R (2001), “Superresolution
HRR ATR with High definition Vector Imaging”, IEEE Transactions on Aerospace and
[8] Liao, X., Runkle, P., and Carin, L. (2002), “Identification of Ground targets From
[9] Mitchell, R.A. and Westerkamp, J.J. (1997), “A statistical Feature based Classifier
For Robust High range Resolution Radar target Identification”, Automatic Target
[11] Shaw, A. K. and Bhatnagar, V. (1998), “Automatic Target Recognition using Eigen
142
[12] Shaw, A K. and Vashist, R. (2000), “HRR-ATR using Eigen-Templates and
[13] Hawley, R.W. (1999), “Complex HRR signal generation for TRUMPETS”,
[15] Dudgeon, D.W. and Lacoss, R.T. (1993), “An overview of automatic target
[16] Novak, L.M., Halversen, S.D., Owirka, G.J. and Heitt, M. (1995), “Effects of
[17] Verly, J.G., Delanoy R.L. (1996), “Model-based automatic target recognition (ATR)
systems for forward looking ground based and airborne imaging laser radars”
[18] Curlander, J.C. and McDonough, R.N. (1991), “Synthetic Aperture Radar”, John
143
[19] Goj, W.W. (1993), “Synthetic Aperture Radar and Electronic Warfare”, Artec
[20] Kovaly, J.J. (1976), “Synthetic Aperture Radar”, Artech House, 1976.
[21] Scheler, D.C. (1991), “MTI and Pulsed Doppler Radar”, Artech House, MA, 1991.
[22] Wehner, D.R. (1987), “High Resolution Radar”, Artech House, MA, 1987.
[23] Scheler, D.C. (1980), “MTI Radar”, Artech House, INC., 1980.
[24] Scheler, D.C. (1991), “MTI and Pulsed Doppler Radar”, AIL Systems, INC., 1991.
[25] Povejsil, D., Raven, R. And Waterman, R. (1961), “Airborne Radar”, D. Van
[27] Kovaly, J.J. (1976), “Synthetic Aperture Radar”, Artech House, 1976.
[28] Hovanessian, S.A. (1980), “Introduction to Synthetic Array and Imaging Radars”.
144
[29] Anderson, J.A., Mohanty, N.C. and Bhattacharya, P.K. (1995), “Model for detection
& tracking of multiple targets”, SPIE PROCEEDINDS, 2984(4): 147-153, April 1995.
[30] Morgan, D.R. (1995), “Automatic target recognition via classical detection theory”,
[31] Jacobs, S.P. and O'Sullivan, J.A. (1997), “High resolution radar models for joint
[32] Xun, Z., Zhaowen, Z. and Guirong, G. (1997), “Automatic HRR target recognition
[33] Yiding, W., Xiangguo, L. and Yunhong, W. (1997), “Recognition of high resolution
[34] Dudgeon, D.E., Grimson, W. and Tenney, R. (1995), “Issues in SAR model-based
[35] Morgan, D.R. (1995), “Automatic target recognition via classical detection theory”,
145
[36] Verly, J.G., Delanoy, R.L. and Lazott, C.H. (1993), “Principles and evaluation of an
automatic target recognition system for synthetic aperture radar imagery based on the use
[39] Chui, C.K. (1992), “An Introduction to Wavelets”, Academic Press, CA, 1992.
[40] Editor Chui, C.K. (1992), “Wavelets: A tutorial in Theory and Applications”,
[41] Daubechies, I. (1990), “The wavelet transform, time frequency localization and
[42] Antonini, M., Barlaud, M. and Mathieu, P. (1990) “Image coding using vector
146
[43] Mallat, S. (1989), “Multifrequency channel decomposition of images and wavelet
models”, IEEE Transactions on Acoustics Speech and Signal Processing, 37: 2091-2110,
1989.
representation”, IEEE Trans. on Pattern Anal. And Machine Intell, 11: 674-693, 1989.
[45] Guo, H., Odegard, J.E., Lang, M., Gopinath, R.A., Selesnick, I.W. and Burrus, C.S.
(1994), “Wavelet based speckle reduction with applications to SAR based ATR”, Proc.,
[46] Pederson, J.F., Saff, E.B. and Schweiker, K. (1993), “Wavelets for training a neural
[47] Tagliarini, G.A., Page, E.W., Kesden, G.M., Chiang, D. and McPartland P.J. (1997),
[48] Bottcher, C. and Strayer, M.R. (1993), “Classical scattering theory of waves from
the view point of an eigenvalue problem and application to target identification”, SPIE
147
[49] Paul, D.B. (1990), “Speech recognition using hidden markov models”, The Lincoln
[50] Rabiner, L.R., Levinson, S.E. and Sondhi, M.M. (1983), “An Introduction to the
[51] Flake, L., Krishnamurthy, A. and Ahalt, S. (1995), “Multi-aperture SAR target
detection using hidden markov model”, SPANN Technical Report TR-95-02, The Ohio
[52] Stroud, P.D. and Gordon, R.C. (1997), “Automated military unit identification in
[54] Jacobs, S.B., O'Sullivan, J.A. and Bekker, Z. (1997), “Recognition of ground targets
[55] Jacobs, S.P., O'Sullivan, J.A., Faisal, M. and Snyder, D.L. (1996), “Automatic target
recognition using high resolution radar”, SPIE Proceedings, The International Society of
148
[56] Worrell, S. (1996), “Air to ground automatic target recognition”, An unclassified
report, 1996.
[57] Stewart, C., Larson, V. and Halsey, J.D. (1992), “Comparison of classification
April 1992.
[58] Bhatnagar, V., Shaw, A.K. and Williams, R. (1998), “Improved Automatic Target
Acoustics, Speech and Signal Processing, Seattle, WA, vol. 5, pp. 2717-2720, May 1998.
[59] Bhatnagar V. (1998), Automatic Target Recognition using High Range Resolution
[60] Forney, G.D., “The Viterbi algorithm”, Proc. IEEE, 1973, 61, 268-278.
[61] Linde, Y., Bujo, A. And Gray R., “An algorithm for vector quantizer design ”, IEEE
[62] Rabiner, L.R., “ A tutorial on hidden Markov models and selected applications in
149
[63] Do M.N. and Wagner M. (1998), “Speaker recognition with small training
[64] Mitra, A. K., Lewis, T. L., Paul, A. S. and Shaw A. K. (2002), “Ultra-Wideband
Radar Data Models and Target Detection with Adaptive Rank Order Filters”, Synthetic
[65] Mitra, A.K. and Lewis, T.M. (2002), “Spectral Analysis of high-frequency SAR
data”, Synthetic Aperture Radar Imagery V: Proceedings of SPIE, Orlando, FL, April
2002.
[66] Chiang, H.C., Moses, R. L. and Potter, L.C. (2000): “Model-based Classification of
Radar Images,” IEEE Transactions on Information Theory, vol. 46, no. 5, pp. 1842 –
[68] Soumekh, M. (1999): “Synthetic Aperture Radar Signal Processing with MATLAB
150
[69] Davis, M., Tomlinson, P. and Paul, M., “Technical Challenges in Ultra-Wideband
Radar Development for Target Detection and Terrain Mapping”, Proceedings of the
Environments”, IEEE National Radar Conference, 1998, Dallas, TX, 12-13 May 1998.
151