DYNAMIC STATE ESTIMATION IN

PRESENCE OF UNCERTAINTY

By
Arpita Mukherjee

Under the Guidance of

Dr. Aparajita Sengupta

RESEARCH THESIS SUBMITTED FOR FULFILLMENT OF THE

REQUIREMENT FOR THE AWARD OF THE DEGREE OF
DOCTOR OF PHILOSOPHY
IN
ELECTRICAL ENGINEERING

BENGAL ENGINEERING AND SCIENCE UNIVERSITY

SHIBPUR, HOWRAH-711103
OCTOBER 2009
Dedicated to
My Family

i
.

ii
Table of Contents

Table of Contents iii

List of Tables vii

List of Figures ix

Abstract xiii

Acknowledgements xv

1 Introduction 1
1.1 Brief Overview of Background Theory . . . . . . . . . . . . . . . . . . 4
1.1.1 Recursive Bayesian Estimation . . . . . . . . . . . . . . . . . 5
1.2 Brief Review on Suboptimal Nonlinear Bayesian Filters . . . . . . . . 8
1.3 State Estimation using Nonlinear Filters . . . . . . . . . . . . . . . . 11
1.3.1 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Brief Overview and Contributions of the Thesis . . . . . . . . . . . . 16

2 State Estimation of Linear Stochastic System Using Recursive Max-

imum a Posteriori Estimator 21
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 MAP Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 The discrete time recursive MAP estimator . . . . . . . . . . . . . . . 24
2.3.1 The algorithm of recursive MAP estimator for a linear system 27
2.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.5 Derivation of Kalman Filter equation from RMAPE . . . . . . . . . . 30
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

iii
3 State Estimation of Ballistic Target Missile from Seeker Measure-
ments using Nonlinear Filters 33
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Mathematical Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2.1 State Space Equations . . . . . . . . . . . . . . . . . . . . . . 42
3.2.2 Measurement Equations . . . . . . . . . . . . . . . . . . . . . 43
3.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

4 Seeker Filter Design of Ballistic Target Missile Incorporating Seeker

Dynamics 51
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.3 Tracking Filter Design incorporating Seeker Model . . . . . . . . . . . 55
4.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5 Kinematic State Estimation of a Ballistic Target Missile from Bad

Seeker Measurements 65
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2 Proposed strategy to handle bad measurement . . . . . . . . . . . . . 67
5.3 The EKF Algorithm implementing New Strategy . . . . . . . . . . . 68
5.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

6 Estimating the Probability Density Function of a Non-Stationary

Non-Gaussian Noise 77
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.2 Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.3 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.4.1 Example I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.4.2 Example II . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.4.3 Example III . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

7 Likelihood Function Modeling of Particle Filter in Presence of Non-

Stationary Non-Gaussian Measurement Noise 95
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

iv
 7.2 SIR Particle Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.3 Likelihood Function Modeling of an SIR Particle Filter . . . . . . . . 100
7.3.1 Algorithm: New SIR particle Filter . . . . . . . . . . . . . . . 101
7.4 Model Order Growth over Iteration for Gaussian Mixture Models of
Non-Gaussian Measurement Noise . . . . . . . . . . . . . . . . . . . . 102
7.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.5.1 Example I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.5.2 Example II . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

8 Nonlinear State Estimation Using Hybrid Robust Unscented Kalman

Filter 117
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
8.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.3 Low Pass Filter for Uncertainty Estimation . . . . . . . . . . . . . . . 120
8.4 Robust Unscented Kalman Filter . . . . . . . . . . . . . . . . . . . . 122
8.5 Calculation of Uncertainty Estimation Error Covariance . . . . . . . . 125
8.6 Hybrid Robust Unscented Kalman Filter . . . . . . . . . . . . . . . . 127
8.7 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
8.7.1 Example I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
8.7.2 Example II . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

9 Conclusion 139
9.1 Future Scope of the Work . . . . . . . . . . . . . . . . . . . . . . . . 141

A Extended Kalman Filter 143

B Unscented Kalman Filter 145

C Particle Filter 149

Bibliography 153

v
.

vi
List of Tables

1.1 Relative computational load . . . . . . . . . . . . . . . . . . . . . . . 12

vii
.

viii
List of Figures

1.1 Case 1: RMS estimation error using EKF, UKF and SIR PF in presence
of Gaussian measurement noise . . . . . . . . . . . . . . . . . . . . . 13
1.2 Case 1: Probability density of Gaussian measurement noise . . . . . . 13
1.3 Case 2: RMS estimation error using EKF, UKF and SIR PF in presence
of non-Gaussian measurement noise . . . . . . . . . . . . . . . . . . . 14
1.4 Case 2: Probability density of non-Gaussian measurement noise . . . 14

2.1 Estimation error of x1 using RMAPE(blue) and KF(red) w.r.t. time

sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2 Estimation error of x2 using RMAPE(blue) and KF(red) w.r.t. time
sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.1 Axes convention of LOS frame w.r.t inertial frame . . . . . . . . . . . 37

3.2 Rotation sequence used to obtain LOS frame from inertial frame . . . 38
3.3 Axes Convention of interceptor body and fin frame . . . . . . . . . . 41
3.4 Interceptor findash and seeker frame . . . . . . . . . . . . . . . . . . 41
3.5 Noise in range rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.6 Noise in gimbal angle along y-axis . . . . . . . . . . . . . . . . . . . . 45
3.7 Noise in gimbal angle along z-axis . . . . . . . . . . . . . . . . . . . . 45
3.8 Noise in LOS rate along y-axis . . . . . . . . . . . . . . . . . . . . . . 46
3.9 Noise in LOS rate along y-axis . . . . . . . . . . . . . . . . . . . . . . 46
3.10 Comparison of estimation error of relative position along X-axis using
EKF, UKF and PF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

ix
3.11 Comparison of estimation error of relative position along Y-axis using
EKF, UKF and PF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.12 Comparison of estimation error of relative position along Z-axis using
EKF, UKF and PF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.13 Comparison of estimation error of relative velocity along X-axis using
EKF, UKF and PF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.14 Comparison of estimation error of relative velocity along Y-axis using
EKF, UKF and PF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.15 Comparison of estimation error of relative velocity along Z-axis using
EKF, UKF and PF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

4.1 2 nd order transfer function . . . . . . . . . . . . . . . . . . . . . . . 54

4.2 Block diagram of seeker model . . . . . . . . . . . . . . . . . . . . . . 56
4.3 Comparison of absolute value of estimation error of relative position
along X-axis with and without incorporating seeker model . . . . . . 61
4.4 Comparison of absolute value of estimation error of relative position
along Y-axis with and without incorporating seeker model . . . . . . 61
4.5 Comparison of absolute value of estimation error of relative position
along Z-axis with and without incorporating seeker model . . . . . . 62
4.6 Comparison of absolute value of estimation error of relative velocity
along X-axis with and without incorporating seeker model . . . . . . 62
4.7 Comparison of absolute value of estimation error of relative velocity
along Y-axis with and without incorporating seeker model . . . . . . 63
4.8 Comparison of absolute value of estimation error of relative velocity
along Z-axis with and without incorporating seeker model . . . . . . 63

5.1 Noise in range rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

5.2 Noise in gimbal angle along y-axis . . . . . . . . . . . . . . . . . . . . 71
5.3 Noise in gimbal angle along z-axis . . . . . . . . . . . . . . . . . . . . 71
5.4 Noise in LOS rate along y-axis . . . . . . . . . . . . . . . . . . . . . . 72

x
5.5 Noise in LOS rate along z-axis . . . . . . . . . . . . . . . . . . . . . . 72
5.6 Estimation error in ∆x using EKF and modified EKF . . . . . . . . . 73
5.7 Estimation error in ∆y using EKF and modified EKF) . . . . . . . . 73
5.8 Estimation error in ∆z using EKF and modified EKF . . . . . . . . . 74
5.9 Estimation error in ∆vx using EKF and modified EKF . . . . . . . . 74
5.10 Estimation error in ∆vy using EKF and modified EKF . . . . . . . . 75
5.11 Estimation error in ∆vz using EKF and modified EKF . . . . . . . . 75

6.1 Estimated PDF at 5th time step . . . . . . . . . . . . . . . . . . . . . 87

6.2 Estimated PDF at 25th time step . . . . . . . . . . . . . . . . . . . . 88
6.3 Estimated PDF at 45th time step . . . . . . . . . . . . . . . . . . . . 88
6.4 Plot of non-stationary glint noise . . . . . . . . . . . . . . . . . . . . 89
6.5 Estimated PDF at 5th time step . . . . . . . . . . . . . . . . . . . . . 90
6.6 Estimated PDF at 25th time step . . . . . . . . . . . . . . . . . . . . 91
6.7 Estimated PDF at 45th time step . . . . . . . . . . . . . . . . . . . . 91
6.8 Histogram of bi-modal distributed data . . . . . . . . . . . . . . . . . 92
6.9 Estimated PDF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

7.1 RMSE of 100 Monte Carlo runs . . . . . . . . . . . . . . . . . . . . . 106

7.2 Average of MSE over time samples with number of particles 50, 100
and 500 (for 100 Monte Carlo runs). . . . . . . . . . . . . . . . . . . 107
7.3 Estimated PDF of student t distribution . . . . . . . . . . . . . . . . 110
7.4 RMSE of 100 Monte Carlo runs . . . . . . . . . . . . . . . . . . . . . 112
7.5 RMSE of 100 Monte Carlo runs . . . . . . . . . . . . . . . . . . . . . 112
7.6 RMSE of 100 Monte Carlo runs . . . . . . . . . . . . . . . . . . . . . 113
7.7 RMSE of 100 Monte Carlo runs . . . . . . . . . . . . . . . . . . . . . 113
7.8 Comparison of average MSE of velocity along x-axis and y-axis for the
number of particles: 5000 - 1000 - 500 - 100 . . . . . . . . . . . . . . 114
7.9 Comparison of average MSE of position along x-axis and y-axis for the
number of particles: 5000 - 1000 - 500 - 100 . . . . . . . . . . . . . . 114

xi
8.1 RMS error along x-axis with unknown ballistic coefficient and aerody-
namic forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
8.2 RMS error along y-axis with unknown ballistic coefficient and aerody-
namic forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8.3 RMS error along x-axis with known ballistic coefficient and aerody-
namic forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8.4 RMS error along y-axis with known ballistic coefficient and aerody-
namic forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.5 RMS estimation error of x2 with unknown input . . . . . . . . . . . . 136
8.6 RMS estimation error of x2 with known input . . . . . . . . . . . . . 136
8.7 RMS estimation error of x2 with unknown input with a larger value of
Qk in case of UKF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

xii
Abstract

The objective of this thesis is state estimation of dynamical systems using the noisy
measurements in presence of uncertainties. First of all a recursive maximum a pos-
teriori estimator (RMAPE) has been derived from the basic theory of ‘Maximum
a posteriori’ parameter estimator for real time state estimation of a linear dynam-
ical system in presence of zero mean white Gaussian noise. It is also shown from
mathematical derivation and simulation results that the RMAPE and the well known
Kalman filter (which is based on ‘minimum mean square error’ estimator) are same.
The rest of the thesis deals with state estimation of nonlinear systems in presence of
different types of uncertainties and can be divided in two parts. In part I, kinematic
state estimation of a ballistic target missile w.r.t. homing (interceptor) missile using
RF seeker (mounted on interceptor missile) measurements in exo-atmospheric region
has been considered. For this problem, comparing the performances of the Extended
Kalman filter (EKF), Unscented Kalman filter (UKF) and particle filter (PF), the
EKF is found to give the best performance regarding estimation error and computa-
tion load. The seeker measurements contain different forms of uncertainties e.g. lag,
jumps, spikes, discontinuity and data loss over and above the usual zero mean white
Gaussian noise. The seeker dynamics has been modeled and included in the system
model of the filter design to handle the lag present in the seeker measurements due to
the mechanism of data collection and transmission to the estimator during real time
state estimation. Next, the EKF algorithm is modified to detect the uncertainties
(jumps, spikes, discontinuity and data loss over) in the measurements and it tunes the
value of measurement noise covariance (R) accordingly. This modified EKF with a

xiii
xiv

time varying noise covariance gives much better results compared to the conventional
EKF. In part II, some generalized techniques are proposed to deal with measurements
and system uncertainties of nonlinear system. Normally the filters assume the noise
present in the state space and measurement model to be zero mean white Gaussian
which is not always true in real world situation. The performance of conventional
filters degrade in presence of non-Gaussian noise. So modeling of non-Gaussian non-
stationary noise has been taken into consideration. A generalized algorithm has been
developed to estimate the parameters of a stationary/non-stationary, Gaussian/non-
Gaussian, zero mean/non-zero mean and uni-modal/multi-modal distributed noise
using Gaussian mixture maximum likelihood approach. The simulation results show
that the estimated noise probability density function (PDF) closely approximates the
actual non-stationary non-Gaussian multi-modal noise PDF. And this new algorithm
has been implemented in a sequential importance resampling (SIR) particle filter for
state estimation of nonlinear system in presence of non-stationary non-Gaussian, non-
zero mean measurement noise. A filter performance further degrades when there is
unknown uncertainties present in the state space model of the system. The uncer-
tainty may be in the form of unknown input, parameter uncertainties, multiplicative
noise, unknown disturbances etc. Finally a robust UKF has been proposed to handle
these uncertainties, also a hybrid robust UKF has been developed to get favorable per-
formance regarding estimation error both in presence and absence of unknown/known,
time-varying/timeinvariant uncertainty.
Acknowledgements

This thesis is the stepping stone of my journey into the vast and ever expanding
horizon of research and development. Thus this thesis forms one the most significant
part of my career in research and development. During the period of the research
that culminated into this thesis, I had the good fortune to meet and interact with
several persons who have provided me with help, support and inspiration during this
entire period. I would like to take this opportunity to tell them a sincere ”thank you”.

Destiny was very kind to me that I got the opportunity to work for my doc-
toral thesis under the guidance and supervision of Dr. Aparajita Sengupta, Assistant
Professor, Department of Electrical Engineering, Bengal Engineering and Science Uni-
versity (BESU), Shibpur. I would like to express my sincerest gratitude to her for her
insightful guidance, constant care, concern, immense encouragement. I am grateful
to her for understanding me and giving me the freedom to progress in my own pace,
which was key to develop my capability as a researcher and thinker.

I owe my deepest gratitude to late Prof. P K Nandi, Professor, Department of

Electrical Engineering, BESU, Shibpur. I was delighted to interact with him and
it was an honour for me to have his expert guidance. He shared his deep and intri-
cate insights with me which helped me to grasp the complexity of my research domain.

I am grateful to the Head of the Department of Electrical Engineering, Prof. S K

Mallik for his cooperation. I want to specially thank Prof. J. Pal, Prof. J. K. Das,
Prof. A. Chakraborty, Prof. M. Sengupta, Prof. G. Bandyopadhyay, Prof. P Syam
and all other faculty members of Department of Electrical Engineering for their help
and support whenever I approached them. I also want to thank the staff members of
this University.

xv
xvi

I wish to thank the Defense Research and Development Laboratory (DRDL), Hy-
derabad for their funding of the project of which I was a member. I am fortunate to
be part of this challenging project and to have the opportunity to interact and learn
from many eminent scientists and professors. I want to thank Dr. R N Bhattachaya,
Dr. Abhijit Bhattacharya, Mr. Vamsi K. Reddy, Scientists of DRDL for very useful
interactions which helped to enrich my conception in my field. I am indebted to Prof.
S. Mukhopadhyay, Professor, Department of Electrical Engineering, IIT Kharagpur,
Prof. T K Ghosal, Professor, Department of Electrical Engineering, Jadavpur Uni-
versity, Kolkata and all the other members of of the project team for their suggestions
and cooperation.

I sincerely acknowledge the cooperation I received from all my present and pre-
vious members of the Control Systems Laboratory Shubhankar, Rajoshri, Sourav,
Maharana-da, Arko, Shruti throughout the period of this work. I specially thank
Arnab for help and support whenever I encountered any problem in software han-
dling.

My deepest gratitude goes to my family. I want to thank my parents Shri Dilip

Kumar Chatterjee, Smt. Anita Chatterjee, my parents-in-law Dr. Narayan Prasad
Mukherjee, Smt. Laxmi Mukherjee for their endless love, support, encouragement
and blessings. I thank my sister Aditi, sister-in-law Monobina and her husband Am-
barish for their love and support. I want to thank my husband Debaprasad Mukherjee
for his valuable suggestions and continuous encouragement.

There are several other individuals without whose presence and support, explicit
or implicit, this thesis would have taken much longer than it has taken. Their names
have not been mentioned here but I thank them from the core of my heart. I seek
their blessings and good wishes beyond this thesis, to my journey into research.

Shibpur, Howrah Arpita Mukherjee

October 2009
Chapter 1

Introduction

Estimation of states of dynamic systems in real time is an essential component in

the design of critical state-of-the-art systems (e.g. target tracking navigation, image

processing, communication channels and networks, biomedical systems, control of

complex systems, distributed computer systems etc). A major problem in estimating

the evolving states of the systems is the presence of uncertainty. Dynamic state

estimation in the presence of uncertainty is a challenging area in various engineering

domains such as aerospace, robotics, bio-medicine, economics, weather prediction,

nuclear engineering etc. Uncertainty in the system or its model arises primarily due

to lack of a complete knowledge about all the system components or the interactions

among various components of systems. The origin of this uncertainty can thus be

either internal or external to the system under consideration or mostly both. In all

these problems, the common factor that throws up the challenge, as of the present

state of the art of these technologies, is the type and degree of uncertainty in the whole

system. This uncertainty can be visualized as some form of ‘noise’, or may be modeled

as uni or multidimensional signals with underlying randomness, and incorporated in

the mathematical and/or computational model of the systems.

1
2

For several years, there have been efforts to design and develop mathematical

and computational models and algorithms for accurate, precise and optimal state

estimation in the presence of different classes of noise incorporated within the models.

The approaches to this problem have relied on Bayesian models, Markov processes,

Monte Carlo, neural networks, genetic algorithms, higher order statistics, information

theory etc. All these techniques have attained some degree of success corresponding to

different classes of problems or real world applications. Sometimes some model and/or

algorithm has been able to provide a precise and optimal solution for a general class

of problems while not being suitable for several other classes of problems. At other

times some model and/or algorithm has performed excellently for a specific particular

engineering real world application while giving unacceptable performance for several

other unrelated real world applications.

One major section of research and development in this area has been the mod-

eling and the filtering of the noise to get more accurate estimation. In one class of

problems the measurement of the states or some related function of the states of the

dynamically evolving system is performed which is then used to estimate the states

of the system. There may also be uncertainty in the sequence of the measurements

made on the system and this is thus termed as noisy measurements. For many prob-

lems an estimate is required every time a measurement is received. In this case a

recursive filter is a convenient solution. A recursive approach means that recursive

data can be processed sequentially rather than as a batch, so that it is not necessary

to store the complete data set nor to reprocess existing data if a new measurement

is available. Such a filter consists of essentially two stages: prediction and update.

The prediction stage uses the system model to predict the state probability density
 3

function (PDF) forward from one measurement time to the next. Since the state

is usually subject to unknown disturbances (modeled as a random noise), prediction

generally translates, deforms and broadens the state PDF. The update operation uses

the latest measurement to modify (typically to tighten) the prediction PDF. This is

achieved using Bayes theorem, which is the mechanism for updating knowledge about

the target state in the light of extra information from new data. In this class of prob-

lems, to analyze and make inferences about a dynamic system, at least two models

are required: first, a model describing the evolution of the state with time (the sys-

tem or dynamic model), and second, a model relating the noisy measurements to the

state (the measurement model). We shall assume that these models are available in a

probabilistic form. The probabilistic state-space formulation and the requirement for

the updating information on receipt of new measurements are ideally suited for the

Bayesian approach [17]-[20]. This provides a rigorous general framework for dynamic

state estimation problems and has been discussed in the Sec 1.1.1.

In these class of problems,both the model and the algorithm of filtering the noise

is critical for the accurate and precision estimation of the states. There have been

several algorithms which attempt to tackle this problem (which have been given a

formal and factual treatment in the following sections of this chapter). The degree

of success of these algorithms have widely varied historically, primarily due to the

parallel evolution of technology and thus requirements for new algorithms. This is

especially true in the area of aerospace engineering and robotics, where there is a

requirement of tracking targets in complex uncertain environments. In this class

of problems, a ‘tracker’ tracks a ‘target’, where the target changes its mechanical

trajectory either due to external influence or due to internal control. The job of the
4

tracker, to track the target successfully, is difficult and prone to failure. This difficulty

must be removed in mission critical engineering applications like missile tracking,

robotic surgery etc. The measurement, subsequent update and finally estimation of

the states of the target along with tracker has to overcome stochastic components

in their mathematical models, either due to inherent noise present within the whole

systems or due to insufficient information incorporated in the mathematical model.

The work embodied in this thesis tries to highlight some of these intricate problems

in the field of dynamic state estimation in the presence of uncertainty along with noise

modeling and designs of filter algorithms, and finally their applications to mechanical

target tracking.

The overview and survey of literature on the basic background theory of ‘dynamic

state estimation in presence of uncertainty’ is outlined in the remaining part of this

chapter. Further detailed introductory background theory and literature survey on

each of the problems (under consideration in this thesis) have been provided in each

of their corresponding chapters. The chapter ends with a comparison of performance

of some of the most popular nonlinear filters, for state estimation of a highly nonlinear

system, and a summary of the contribution of this thesis.

1.1 Brief Overview of Background Theory

Let us start with the definition of “estimation”. We can say that estimation is the

process of inferring the values of variables that characterize the system. These vari-

ables are classified as either the parameter (a time invariant quantity) or the state

(time variant quantity) of the system. This leads to two fundamental forms of estima-

tion - parameter estimation and state estimation. The basic techniques of parameter
 5

estimation are maximum likelihood (ML) estimate, maximum a posterior (MAP)

estimate, least square (LS) estimate and minimum mean square error (MMSE) es-

timate. These are some of the basic parameter estimation techniques that provide

the foundation of state estimation. The states of a system are estimated from noisy

measurement data fed back by the sensor component of the system. When the system

and observation models are linear and the noise is modeled as Gaussian, the Kalman

filter [15, 16](a “minimum mean square error” filter) is the optimal filter. However

in most of the cases the system dynamics and observation equations are nonlinear.

Thus nonlinear filtering has been the focus of interest in statistical and engineering

community for more than 30 years, the foundation of which is based on Bayesian

probabilistic approaches (Sec. 1.1.1).

1.1.1 Recursive Bayesian Estimation

In the Bayesian approach to dynamic state estimation one attempts to construct the

posterior probability density function (PDF) of the state, based on overall available

information, including the sequence of received measurements. If either the system

or measurement model is nonlinear, the posterior PDF will be non Gaussian. Since

this PDF embodies all available statistical information, it may be regarded to be the

complete solution to the estimation problem. In principle, an optimal (with respect

to any criterion) estimate of the state may be obtained from the posterior PDF.

Let us introduce the state vector xk ∈ Rnx , where nx is the dimension of the state

vector; R is a set of real numbers: k ∈ N is the time index: and N is the set of

natural numbers. Here index k is assigned to a continuous time instant tk , and the

sampling interval Tk−1 = tk − tk−1 may be time dependent (i.e. a function of k). The
6

state evolves according to the following discrete-time stochastic model:

xk = fk−1 (xk−1 , wk−1 ) (1.1.1)

where fk−1 is a known, possibly nonlinear function of the state xk−1 and wk−1 is

referred to as a process noise sequence. Process noise caters for any unmodeled

effects or unforeseen disturbances in state model. The objective of nonlinear filtering

is to recursively estimate xk from measurements zk ∈ Rnz . The measurements are

related to the system state via measurement equation:

zk = hk (xk , vk ) (1.1.2)

where hk is a known, possibly nonlinear function and vk is a measurement noise

sequence. The noise sequences wk−1 and vk will be assumed to be white, with known

probability density functions and mutually independent. The initial state of the

system is assumed to have a known PDF p(x0 ) and also to be independent of noise

sequences.

We seek filtered estimates of xk based on the sequence of all available measure-

ments Zk = (zi , i = 1....k) up to time k. Thus, it is required to construct the posterior

PDF p(xk |Zk ). The initial density of the state vector is p(x0 ) = p(x0 |z0 ), where z0 is

the set of no measurements. Then, in principle, the PDF p(xk |Zk ) may be obtained

recursively in two stages: prediction and update.

Suppose that the required PDF p(xk−1 |Zk−1 ) at time k − 1 is available. The

prediction stage uses the system model (1.1.1) to obtain the prediction density of the

state at time at time k via the Chapman-Kolomogrov equation:

Z
p(xk |Zk−1 ) = p(xk |xk−1 )p(xk−1 |Zk−1 )dxk−1 (1.1.3)
 7

It may be noted that in (1.1.3), use has been made of the fact that p(xk |xk−1 , Zk−1 ) =

p(xk |xk−1 ) as (1.1.1) describes a Markov process of order one [3, 106]. The probabilis-

tic model of state evolution (often referred to as transitional density), p(xk |xk−1 ), is

defined by the system equation (1.1.1) and the known statistics of wk−1 . At time step

k when a measurement zk becomes available, the update stage is carried out. This

involves an update of the prediction (or prior) PDF via the Bayes’ rule:

p(xk |Zk ) = p(xk |zk , Zk−1 )

p(zk |xk , Zk−1 )p(xk |Zk−1 )
=
p(zk |Zk−1 )
p(zk |xk )p(xk |Zk−1 )
=
p(zk |Zk−1 )

where the normalizing constant

Z
p(zk |Zk−1 ) = p(zk |xk )p(xk |Zk−1 )dxk (1.1.4)

depends on the likelihood function p(zk |xk ), defined by the measurement model (1.1.2)

and known statistics of vk . In the update stage (1.1.4), the measurement zk is used to

modify the prior density to obtain the required posterior density of the current state.

The recurrence relations (1.1.3) and (1.1.4) form the basis for optimal Bayesian

solution. Knowledge of the posterior density p(xk |Zk ) enables one to compute an

optimal state estimate with respect to any criterion. For example, the minimum

mean square error (MMSE) estimate is the conditional mean of xk .

Z
x̂M M SE
k|k = E[xk |Zk ] = xk .p(xk |Zk )dxk , (1.1.5)

while the maximum a posteriori (MAP) estimate is the maximum of p(xk |Zk ):

x̂M AP
k|k = arg max p(xk |Zk ) (1.1.6)
xk
8

Similarly, a measure of accuracy of a state estimate (e.g., covariance) may also be

obtained from p(xk |Zk ).

The recursive propagation of the density, given by (1.1.3) and (1.1.4) is only a

conceptual solution in the sense that in general it cannot be determined analytically

in most cases excepting those in which the system and measurement equations are

linear and the noise present is Gaussian in nature. In other general cases, the imple-

mentation of the conceptual solution requires the storage of the entire (non-Gaussian)

PDF which is equivalent to an infinite dimensional vector [1]. Only in a restrictive

set of cases, the posterior density can be exactly and completely characterized by a

sufficient statistics of fixed and finite dimension. Since in most real world situations

the practical solution of (1.1.3) and (1.1.4) is intractable, one has to use approximates

or suboptimal Bayesian algorithms [2].

1.2 Brief Review on Suboptimal Nonlinear Bayesian

Filters

A huge number of approximate nonlinear filters have been proposed [4] - [5], [16].

Some of them are fairly general and others are more specific to a particular application.

Here only a brief discussion is given of those nonlinear filters that have been adopted

by a wider scientific and engineering community. These filters can be grouped into

five broad classes:

1. Analytic approximations: The main feature of these filters are that they approx-

imate (linearize) the nonlinear function of the state dynamic and measurement

models [6]-[15].

The extended Kalman filter (EKF) (Appendix A), the higher-order EKF and
 9

iterated EKF are referred to as analytic approximations because the Jacobians

of the nonlinear functions fk and hk have to be worked out analytically.

The EKF always approximates p(xk |Zk ) to be Gaussian so if the models (1.1.1)

and (1.1.2) are severely nonlinear then the non-Gaussianity of the true posterior

density will be more pronounced. In such case the performance of the EKF will

be degraded significantly.

2. Numerical approximations: This class of non-linear filters perform numerical in-

tegration in order to solve the multidimensional integral in (1.1.3) and (1.1.4).

Numerical methods are also referred to as approximate grid-based methods and

essentially they replaced the integral by summation involving discretization of

integration variable [4], [7]-[9].

Hidden Markov model (HMM) filters [10], [11] are an application of such ap-

proximate grid based methods in a fixed-interval smoothing context and have

been used extensively in speech processing

The grid must be sufficiently dense to get good approximation to the continuous

state space. The grid-based methods overcome the limitation by approximating

the entire posterior density. But this methods have a serious drawback that

the computational complexity increases exponentially with respect to the state

vector dimension. This disadvantage prevents the widespread application of

this methods.

3. Gaussian sum or multiple model filters: The key idea of the Gaussian sum filter

[12] is to approximate the required posterior density p(xk |Zk ) by a Gaussian

mixture (a weighted sum of Gaussian density functions).

10

The static multiple model (MM) estimator and dynamic MM [15] are the pop-

ular Gaussian sum filter for tracking maneuvering target. The interacting mul-

tiple model (IMM) [15] algorithm is the most popular and widely used dynamic

MM estimator for maneuvering target tracking. However, there is a lack of an

automatic procedure for the computation of weights, means and covariances of

the Gaussian sum filter for a general nonlinear filtering problem.

4. sampling approaches: The central idea is to represent the posterior PDF by a

set of deterministically chosen samples and associated weights. The unscented

Kalman filter (UKF) can be grouped in this category. The UKF [30]-[33] has

been described briefly in Appendix B.

5. Monte Carlo approximation: It is a technique for implementing a recursive

Bayesian filter by Monte Carlo simulation. The key idea is to represent the

required density function by a set of random samples or particles with associated

weights and to compute estimates based on these samples/particles and weights.

These filters are termed as particle filter and are often similar to importance

sampling methods[17]-[20]. There are several version of particle filters, some of

them are

i) sampling importance resampling (SIR) filters [21] (SIR filter briefly dis-

cussed in Appendix C),

ii) Auxiliary SIR filters [23]:It is a variant of the standard SIR. The basic idea

is to perform resampling step at time k − 1 (using the available measure-

ment at time k), before the particle are propagated in time k.

iii) Particle filters with an improved sample diversity: There are mainly two
 11

techniques to improve diversity among the particles. The first one is the

regularized particle filter (RPF) [24] and the second one is the Markov

chain Monte Carlo move step [25].

iv) Local linearization particle filters [26], [27], [102], [101]: The optimal im-

portance density can be approximated by incorporating the most current

measurement zk via a bank of extended or unscented Kalman filters.

v) Multiple-Model particle filter [28]- [29]: This filter is proposed to perform

nonlinear filtering with switching dynamics.

1.3 State Estimation using Nonlinear Filters

Among the Bayesian nonlinear filters mentioned above some of the popular and most

widely used suboptimal nonlinear filters are - EKF [13]-[16], UKF [30]-[33] and SIR

PF [17] - [20].

The performance of these three filters have been compared regarding estimation

error and computational load for a standard nonlinear system in the presence of both

Gaussian and non-Gaussian (heavy tailed PDF) measurement noise separately. Brief

algorithms of EKF, UKF and SIR PF are given in Appendix A, B and C respectively.

1.3.1 Simulation Results

Here the state of a nonlinear system [18, 21] has been estimated in presence both

Gaussian and non-Gaussian measurement noise. The results using the three filters,

EKF, UKF and SIR PF are shown. The system is given by,
25xk−1
xk = 0.5xk−1 + + 8cos(1.2(k − 1)) + wk (1.3.1)
[1 + xk−1 ]2
x2k
yk = + vk (1.3.2)
20
12

Table 1.1: Relative computational load

EKF UKF SIR PF

1 3 500

where wk is a zero mean Gaussian white noise with variance 10. Two different cases

have been considered - in case 1, vk is a zero mean Gaussian white noise with variance

1 and in case 2, vk is a zero mean non-Gaussian (skewed PDF) noise (simulated by

student t distribution with parameter - degree of freedom = 2). The PDF of the

measurement noise for both the cases are shown in Figs. 1.2 and 1.4 respectively.

The initial mean and covariance of the state are assumed as, x0 = 0.1 and P0 = 5

respectively, for all the filters. The process noise and measurement noise covariances

for all the filters have been taken as Q = 10 and R = 1 respectively for both the cases

(case 1 and case 2). The number of samples or particles has been chosen to be 500

in case of the SIR PF. The root mean square (RMS) state estimation error for 200

Monte Carlo runs have been plotted for cases 1 and 2 in Figs. 1.1 1.3 respectively.

The computational load of these filter algorithms are compared by measuring the time

spent to execute the MATLAB code for each filter. Comparison of the computation

loads is given in Table 1.1. This table shows the time spent to execute the MATLAB

code of each filter relative to that by EKF.

1.3.2 Discussion

The simulation study compares the performance of the three nonlinear filters (EKF,

UKF and SIR PF). The filters estimate the states of a standard nonlinear system in

presence of Gaussian (case 1) and non-Gaussian noise (case 2). The RMS estimation
 13

50
PF
45 EKF
UKF

State estimation RMS error

40

35

30

25

20

15

10

0
0 10 20 30 40 50
Time

Figure 1.1: Case 1: RMS estimation error using EKF, UKF and SIR PF in presence
of Gaussian measurement noise

PDF of Measurement noise

0.4

0.35

0.3
Probability density

0.25

0.2

0.15

0.1

0.05

0
−3 −2 −1 0 1 2 3 4
PDF argument

Figure 1.2: Case 1: Probability density of Gaussian measurement noise

14

70
PF
EKF
60 UKF
State estimation RMS error

50

40

30

20

10

0
0 10 20 30 40 50
Time

Figure 1.3: Case 2: RMS estimation error using EKF, UKF and SIR PF in presence
of non-Gaussian measurement noise

PDF of Measurement noise

0.4

0.35

0.3
Probability density

0.25

0.2

0.15

0.1

0.05

0
−15 −10 −5 0 5 10 15 20 25
PDF argument

Figure 1.4: Case 2: Probability density of non-Gaussian measurement noise

 15

error of 200 Monte Carlo runs shows that in case 1, where the measurement noise

is Gaussian, the error of EKF is much higher compared to the same for UKF and

SIR PF. Since the system used in the example is highly nonlinear, the error due

to linearization degrades the performance of the EKF. The RMS estimation error

is lowest in case of SIR PF while the performance of UKF is comparable with SIR

PF. Both EKF and UKF work on the assumption of a Gaussian PDF. Though the

noise in the system is Gaussian in case 1, but due to the nonlinearity in the system

it becomes non-Gaussian. So the difference in the performance between the SIR PF

and the UKF appears to be due to the Gaussian assumption on noise. In case of the

EKF the error is two fold - due to linearization as well as Gaussian assumption. Now

to aggravate the situation further (case 2), a non-Gaussian measurement noise has

been introduced (Fig 1.4) in case 2. Finally the RMS estimation error of 200 Monte

Carlo runs has been plotted. In this case the estimation error of EKF and UKF is

higher compared to the that in case 1. So it appears that the increase is due to an

increase in the non-Gaussian content of the noise in the system. But the performance

of SIR PF in both the cases is comparable. The results thus show that the SIR PF

gives best performance compared to the EKF and UKF for a highly non-linear system

and in presence of non-Gaussian noise. The estimation accuracy of SIR PF depends

on the number of particles which is in turn proportional to the computational load.

The computational load of SIR PF is however much higher compared to EKF and

UKF - comparison of the computational load is shown in Table 1.1 for this example.

So the final choice of the filter to be employed for state estimation of nonlinear and

non-Gaussian systems is guided by accuracy as well as computational burden.

16

In this chapter the basic background theory of dynamic state estimation in pres-

ence of uncertainty is briefly discussed. Further detailed introductory background the-

ory and literature survey on each of the problems (under consideration in this thesis)

have been provided in each of their corresponding chapters. The thesis is organized

as follows.

1.4 Brief Overview and Contributions of the The-

sis

1. Chapter 2: For state estimation of linear dynamical systems in presence of

zero mean white Gaussian noise, the well known Kalman Filter (KF) is the

optimal minimum mean square error (MMSE) estimator. In this chapter a

recursive maximum a posteriori estimator (RMAPE) has been developed for

state estimation of linear dynamical system with Gaussian assumption [133]-

[134]. It has been derived from the fundamentals of parameter estimation,

namely, the maximum a posteriori (MAP) estimate. Theoretical and simulation

results show that RMAPE and KF may be considered one and the same for

online state estimation of a linear dynamical system in presence of zero mean

white Gaussian noise. There are several applications of Kalman filter (MMSE

estimator) and MAP estimator separately in different fields. So this work will

give a new interpretation of these filters incase of state and parameter estimation

of linear Gaussian systems.

The rest of the work deals with suboptimal nonlinear filtering for real time state esti-

mation in presence of different types of uncertainty in the system and measurement.

Though a vast literature exists in this field, still it remains an important research
 17

problem because of the fact that the performance of these filters degrade in presence

of certain types of uncertainty. This thesis addresses some of these issues. The work

can be divided into two parts. In the part I, the problem of tracking a ballistic mis-

sile has been considered, it being one of the most important and challenging practical

applications of estimation theory. In part II, Some generalized techniques have been

proposed to handle different types of uncertainty in the system and measurement

noise.

Part I

1. Chapter 3: Some of the most popular nonlinear filters (EKF, UKF and SIR

PF) have been used for real time state estimation of a tactical ballistic missile

(target) with respect to a tracking homing missile (interceptor) from noisy radio

frequency (RF) seeker measurements when the engagement takes place in exo-

atmosphere [137]. The seeker measurements have been generated in a 6-degree

of freedom (DOF) simulation environment. The simulation results of the EKF,

UKF and SIR PF have been compared to evaluate their performances. Com-

paring both the estimation error and computation load, EKF is found to be a

better filter compared to UKF and SIR PF for state estimation of a ballistic

target missile from RF seeker measurements in exo-atmosphere.

2. Chapter 4: The real time states of a tactical ballistic missile in exo-atmosphere

are estimated using the measurements from an RF seeker which contains lag

due the mechanism of data collection and transmission to the estimator in the

system. This lag in the measurements degrades the estimator performance as it

faces problems in updating the states in current time step with measurements
18

containing time lag. In this chapter, an extended Kalman filter has been de-

signed to estimate the states of a ballistic target missile by augmenting the

seeker dynamics with the system model of the estimator in order to handle the

lag present in the seeker measurements due to processing and transmitting data

obtained in different reference frames [138].

3. Chapter 5: The seeker measurements which are used for real time state esti-

mation of a ballistic missile with respect to a tracking missile are contaminated

by high degree of noise which is due to eclipsing, glint, radar cross section fluc-

tuation and thermal effect. These noise appear in the form of jumps, spikes,

discontinuity and data loss over and above the usual zero mean white Gaussian

noise present in a radio frequency (RF) seeker. These measurement are termed

here, as ‘bad measurements’. The extended Kalman filter (EKF) algorithm

has been suitably modified to detect the instant when a bad measurement is

received and tune the value of measurement noise covariance (R) accordingly

[135]-[136]. This modified EKF with a time varying noise covariance shows

much better performance compared to conventional EKF.

Part II

4. Chapter 6: In practical situation, the noise present in the model may non-

stationary and non-Gaussian. The performance of these filters degrades in

presence of such type of noise. In this chapter a generalized algorithm has

been proposed to estimate the parameters of noise which may be stationary

or non-stationary, Gaussian or non-Gaussian, zero mean or non-zero mean and

uni-modal or multi-modal distributed noise [139]-[140]. The non-Gaussian noise

is modeled using Gaussian mixture PDFs and the parameters are estimated by
 19

maximizing the log likelihood function. Simulation results demonstrate the util-

ity and validity of the new algorithm. The proposed algorithm is applicable in

practical situations where it is possible to acquire noise data beforehand (at

least through simulation).

5. Chapter 7: If appropriate knowledge of noise present in the measurements is

provided in the filters during state estimation then, it is expected to increase the

filtering efficiency. A generalized algorithm for estimating noise PDF (proposed

in Chapter 6), has been incorporated in the filter design to improve the filter

performance in presence of non-stationary non-Gaussian measurement noise.

The measurement noise is modeled by Gaussian mixture probability density

functions and the noise parameters are estimated by maximizing the log like-

lihood function of the noise model by preprocessing the data history of the

measurement noise. This model is then included in the likelihood function of

the SIR particle filter (PF) at each time step for realtime state estimation of

the system [141]-[142]. In this way the new SIR particle filter is able to estimate

states in presence of any type of white measurement noise sequence.

6. Chapter 8: The state estimation of a nonlinear system in presence of unknown,

unmodeled uncertainty is of a great difficulty. In this chapter a robust non-

linear filter has been proposed to estimate states in presence of any unknown

uncertainty in the state space model. A robust UKF has been developed by

combining a standard UKF with an uncertainty estimator. A low pass filter

is used to estimate the uncertainty present in the system and is included in

the filter design. The proposed robust UKF does not require any the specific

knowledge of the plant uncertainty model. This filter is capable of estimating

20

the states of any nonlinear system in presence of various uncertainty e.g. pa-

rameter uncertainty or unknown input or multiplicative noise or biased process

noise or in the presence of large disturbances etc. Also a hybrid robust UKF has

been proposed which gives favorable performance both in absence and presence

of uncertainty.

The work documented in this thesis, as briefly summarized above, has led to the

research papers [133] to [142].

Chapter 2

State Estimation of Linear

Stochastic System Using Recursive
Maximum a Posteriori Estimator

Kalman Filter (KF) is the optimal state estimator for linear dynamical systems in

presence of zero mean white Gaussian noise. It is conceived to be a minimum mean

square error (MMSE) estimator. In the present work a recursive maximum a poste-

riori estimator (RMAPE) has been developed from the basic parameter estimation

technique - maximum a posteriori (MAP) estimate. Theoretical and simulation re-

sults shows that RMAPE and KF may be considered one and the same for online

state estimation of a linear dynamical system in presence of zero mean white Gaus-

sian noise. It is shown that the KF can be derived from the RMAPE algorithm i.e.

this work also shows a different way to derive KF.

2.1 Introduction

State estimation of a stochastic system is one of the most fundamental problems in

control theory and signal processing. The Kalman filter has been one of the most

widely used techniques applied to recursive estimation problems during the last 50

21
22

years. It is well known that the Kalman filter is the optimal MMSE [13], [15] state

estimator for a linear system under the assumption of Gaussian statistics of process

and measurement noise. An MMSE estimator minimizes the expected value (mean)

of the square of the estimation error conditioned on the measurement. On the other

hand a maximum a posteriori (MAP) estimator estimates the value of the parameter

that maximizes the posterior probability density function (PDF) [15]. The MAP

estimator has been widely advocated in many engineering fields.

In [34], a method was presented for motion compensated image coding based upon

a two-step displacement estimation procedure. The first step utilized a maximum a

posteriori (MAP) estimator to determine the best integer displacement, while the

second step required solving for the regression coefficients that supplied the same

information as the noninteger portion of the displacement. In [35], the MAP was

defined for the estimation of the parameters of discrete-time auto regressive moving

average (ARMA) processes observed in white noise. The author in [36], studied the

performance of a MAP estimator for estimating the time shift of an inhomogeneous

casually filtered Poisson process in the presence of additive Gaussian noise. Exact

distribution of edge-preserving MAP estimator has been derived for linear signal

models with Gaussian measurement noise [37]. A re-weighted interacting multiple

model (IMM) algorithm has been derived by a recursive implementation of a MAP

state sequence estimator for computing the optimal conditional mean state estimate

for a jump Markov linear system [38]. Paper [39] proposed a low complexity iterative

maximum a posterior (MAP) channel estimator for serially concatenated systems

over frequency-nonselective Rayleigh fading channel. In [40], the explicit structure

of the approximate MAP estimator was derived for estimation of the frequency and
 23

phase of a single sinusoidal observed in additive white Gaussian noise, via a priori

probability density function of the unknown frequency and carrier phase. Distributed

algorithms of MAP and linear MMSE schemes were developed for optimal estimation

of stationary random signals and smoothing of dynamical processes based on generally

correlated observations collected by ad hoc wireless sensor networks [41]. Paper [42]

presented a reduced complexity MAP channel estimator with iterative data detection

for orthogonal frequency division multiplexing (OFDM) systems over mobile multiple-

input multiple-output channels. Some other applications of the MAP estimator are -

derivation of exact PDF of MAP estimator [37], and the MAP estimator for motion

compensated image coding [34].

There are thus numerous applications of the MAP estimator in the field of im-

age processing, signal processing, wireless communication and control system. It is

seen that the MAP estimator has been used for parameter estimation [35, 40], delay

parameter estimation [36], channel parameter estimation [39, 42]. The parameter

estimation can be viewed as a static estimation problem [15]. Among these several

applications of the MAP estimator, neither of them have been used for the purpose

of dynamic state estimation except [38], where the MAP state sequence has been

estimated - but that too is done off-line on a batch of data. In this present work,

an algorithm based on the MAP estimate has been developed to recursively esti-

mate the realtime states of a linear dynamical system in the presence of zero mean

white Gaussian noise. The algorithm has been developed by maximizing the posterior

probability density function. It is also shown that the RMAPE equations lead to the

same equations as in the Kalman filter algorithm so this work can be interpreted as

a special way to derive KF algorithm. There are several applications of Kalman filter
24

(MMSE estimator) and the MAP estimator, individually, in the field of engineering.

The derivation and also the simulation results of the present work show that taking

either of the two approaches leads to the same results. This derivation hereby gives

a better insight to the problem of dynamic state estimation of stochastic systems.

2.2 MAP Estimate

The method for estimating of a random parameter by maximization of the posterior

PDF yields the maximum a posteriori (MAP) estimator [15].

x̂M AP (Z) = arg max[p(x|Z)] = arg max[p(Z|x)p(x)] (2.2.1)

x x

where p(·) denotes probability density function.

The maximum a posterior (MAP) estimate of x, depends on the observations Z and

requires an additional prior p(x) on the parameter. This estimator can be utilized

for state estimation of discrete time stochastic systems in the presence of a sequence

of noisy measurements. In the next section a recursive algorithm is developed imple-

menting a MAP estimator for state estimation of a linear stochastic system.

2.3 The discrete time recursive MAP estimator

A recursive MAP estimator has been developed to estimate the states of a discrete

time linear stochastic system from noisy measurements.

The state at the k-th time step is given by:

xk = φk−1 xk−1 + ωk−1 (2.3.1)

The measurement model is:

zk = Hk xk + vk (2.3.2)
 25

where ωk−1 and vk are zero mean white Gaussian noise sequences with covariances

Qk−1 and Rk respectively, φk−1 and Hk are state and measurement transition matrix

respectively.

Bayesian approach has been taken for probabilistic state space formulation and up-

dating of information on receiving a new measurement. In Bayesian dynamic state

estimation, the posteriori probability density function (PDF) is evaluated based on

all available information including the sequence of received measurements. The pos-

teriori PDF, according to the Bayesian approach, may be written as [15]

p(zk |xk )p(xk |z1:k−1 )

p(xk |z1:k ) = (2.3.3)
p(zk |z1:k−1 )

Now, by the definition, the MAP estimate is given by

x̂M
k
AP
= arg max p(xk |z1:k ) (2.3.4)
xk

To characterize the Gaussian conditional density p(xk |z1:k−1 ) completely, its mean

and covariance will now be computed. Let x̂k|k−1 denote the conditional mean of xk

before the measurement zk is taken and processed, i.e.

x̂k|k−1 = E[xk |z1:k−1 ] (2.3.5)

This mean is time propagated using (2.3.1)

x̂k|k−1 = φk−1 x̂k−1|k−1 (2.3.6)

Similarly, defining Pk|k−1 to be the conditional covariance of xk before the measure-

ment zk is taken and processed,

Pk|k−1 = E{[xk − x̂k|k−1 ][xk − x̂k|k−1 ]T |z1:k−1 } (2.3.7)

26

Then the conditional covariance propagates in time according to: (considering xk and

ωk are independent.)

Pk|k−1 = φk−1 E{[xk−1 − x̂k−1|k−1 ][xk−1 − x̂k−1|k−1 ]T |z1:k−1 }φTk−1

T
+ E[ωk−1 ωk−1 ] (2.3.8)

= φk−1 Pk−1|k−1 φTk−1 + Qk−1 (2.3.9)

where Qk−1 is the process noise covariance (assumed symmetric and positive definite).

Finally the conditional density p(xk |z1:k−1 ) we have been seeking can now be written

as

p(xk |z1:k−1 ) = [(2π)n/2 |Pk|k−1 |1/2 ]−1

1 −1
× exp{− (xk − x̂k|k−1 )T Pk|k−1 (xk − x̂k|k−1 )} (2.3.10)
2

n being the dimension of the state vector xk

Now the Gaussian density p(zk |xk ) in (2.3.3) may be completely specified by its mean

and covariance [16].

p(zk |xk ) = [(2π)m/2 |Rk |1/2 ]−1

1
× exp{− (zk − Hk xk )T Rk−1 (zk − Hk xk )} (2.3.11)
2

Finally using (2.3.3), the MAP estimate in (2.3.4) may be evaluated as:

x̂M AP = arg max p(zk |xk )p(xk |z1:k−1 ) (2.3.12)

xk
1 1 −1
or, x̂M AP = arg max exp{− (xk − x̂k|k−1 )T Pk|k−1 (xk − x̂k|k−1 )}
xk Ck 2
1
× exp{− (zk − Hk xk )T Rk−1 (zk − Hk xk )} (2.3.13)
2

where Ck = [(2π)n/2 |Pk|k−1 |1/2 ][(2π)m/2 |Rk |1/2 ].

Using (2.3.12), the final recursive equations of MAP are: (The derivation is included
 27

in Section 2.5, inspired by [43])

−1
x̂k|k = (Pk|k−1 + HkT Rk−1 Hk )−1 (HkT Rk−1 zk + Pk|k−1
−1
x̂k|k−1 ) (2.3.14)
−1
Pk|k = (Pk|k−1 + HkT Rk−1 Hk )−1 (2.3.15)

The RMAPE algorithm is summarized below.

2.3.1 The algorithm of recursive MAP estimator for a linear

system

1. a priori mean

x̂k|k−1 = φk−1 x̂k−1|k−1 (2.3.16)

2. a priori covariance

Pk|k−1 = φk−1 Pk−1|k−1 φTk−1 + Qk−1 (2.3.17)

3. a posteriori mean

−1
x̂k|k = (Pk|k−1 + HkT Rk−1 Hk )−1 (HkT Rk−1 zk + Pk|k−1
−1
x̂k|k−1 ) (2.3.18)

4. a posteriori covariance

−1
Pk|k = (Pk|k−1 + HkT Rk−1 Hk )−1 (2.3.19)

2.4 Simulation Results

The recursive MAP estimator derived here has been used for the state estimation of

a linear dynamical model and the results have been compared with the Kalman filter.

A simple linear two dimensional state model in presence of zero mean white Gaussian
28

noise is taken as a test model.

The state space model is:

" # " #
1 1 0
xk = xk−1 + wk−1 (2.4.1)
0 1 1

wk is zero mean, white," Gassian

# noise with covariance given by Q = 1
x1
The state vector xk =
x2
The observation equation is given by
h i
zk = 1 0 xk + v k (2.4.2)

vk is zero mean, white, Gaussian noise with "covariance,

#R=1
10 0
The initial state covariance is taken as P0 = and the initial state vector is
0 10
" #
0
taken as x0 = . The estimation error of the state variable x1 and x2 is shown
10
in Fig. 2.1 and Fig. 2.2 using both RMAPE and KF. The simulation results show

that estimation error plots using these two estimators overlap. RMAPE and KF give

exactly same results apparently.

 29

Figure 2.1: Estimation error of x1 using RMAPE(blue) and KF(red) w.r.t. time
sample

Figure 2.2: Estimation error of x2 using RMAPE(blue) and KF(red) w.r.t. time
sample
30

2.5 Derivation of Kalman Filter equation from RMAPE

The a posteriori covariance of the MAP estimator is:

from (2.3.19)

−1
Pk|k = (Pk|k−1 + HkT Rk−1 Hk )−1

= Pk|k−1 − Pk|k−1 HkT [Hk Pk|k−1 HkT + Rk ]−1 Hk Pk|k−1 (2.5.1)

using matrix inversion lemma [(P −1 + H T R−1 H)−1 = P − P H T (HP H T + R)−1 HP ]

Define the matrices

Sk = Hk Pk|k−1 HkT + Rk (2.5.2)

Kk = Pk|k−1 HkT Sk−1 (2.5.3)

Using (2.5.2) and (2.5.3), (2.5.1) can be written as:

Pk|k = Pk|k−1 − Kk Hk Pk|k−1 (2.5.4)

or, Pk|k = [I − Kk Hk ]Pk|k−1 (2.5.5)

Now a posteriori mean from (2.3.14) and (2.3.15) is:

−1
x̂k|k = (Pk|k−1 + HkT Rk−1 Hk )−1 (HkT Rk−1 zk + Pk|k−1
−1
x̂k|k−1 )
−1
= Pk|k (HkT Rk−1 zk + Pk|k−1 x̂k|k−1 )
−1
= Pk|k HkT Rk−1 zk + Pk|k Pk|k−1 x̂k|k−1 (2.5.6)

Replacing (2.5.5) in (2.5.6),

x̂k|k = Pk|k HkT Rk−1 zk + [I − Kk Hk ]x̂k|k−1 (2.5.7)

 31

First term in the right hand side of the (2.5.7) can be evaluated using (2.5.1), (2.5.2)

and (2.5.3) as:

Pk|k HkT Rk−1 = {Pk|k−1 HkT − Pk|k−1 HkT [Hk Pk|k−1 HkT + Rk ]−1

× Hk Pk|k−1 HkT }Rk−1

= Pk|k−1 HkT (Hk Pk|k−1 HkT + Rk )−1

× {Hk Pk|k−1 HkT + Rk − Hk Pk|k−1 HkT }Rk−1

= Pk|k−1 HkT (Hk Pk|k−1 HkT + Rk )−1 Rk Rk−1

= Pk|k−1 HkT Sk−1

= Kk (2.5.8)

from (2.5.8) and (2.5.7) we get

x̂k|k = Kk zk + [I − Kk Hk ]x̂k|k−1

x̂k|k = x̂k|k−1 + Kk [zk − Hk x̂k|k−1 ] (2.5.9)

which are same as the update equations for the mean estimate in Kalman filter. The

derived algorithm is given below

1. A priori mean (2.3.16)

x̂k|k−1 = φk−1 x̂k−1|k−1 (2.5.10)

2. a priori covariance (2.3.17)

Pk|k−1 = φk−1 Pk−1|k−1 φTk−1 + Qk−1 (2.5.11)

3. a posteriori mean (2.5.9)

x̂k|k = x̂k|k−1 + Kk [zk − Hk x̂k|k−1 ] (2.5.12)

32

4. a posteriori covariance (2.5.5)

Pk|k = [I − Kk Hk ]Pk|k−1 (2.5.13)

5. gain matrix (2.5.2, 2.5.3)

Kk = Pk|k−1 HkT (Hk Pk|k−1 HkT + Rk )−1 (2.5.14)

The above equations (2.5.10-2.5.14) derived from the recursive MAP estimator also

appear in conventional Kalman filter algorithm [14].

2.6 Conclusion

The simulation results show that the recursive MAP estimator gives exactly similar

results as Kalman filter which is an optimal MMSE estimator. The RMAPE and KF

algorithm will be same as far as the Gaussian assumption is valid. As the KF is based

on minimum mean square error estimator - for random parameter, the mean square

error will be minimum when the estimated value is equal to the mean of the probability

density function (PDF). The RMAPE is based on maximum a posteriori estimator

which maximize the posterior PDF so estimated value will be the mode (or peak) of

the posterior PDF. Now a Gaussian PDF is symmetrical about its mean which is also

the peak value of the PDF. So in case of a Gaussian assumption, the estimated value

for both the cases will be similar. Here the Kalman filter algorithm (2.5.10-2.5.14) is

derived from the recursive MAP estimator algorithm (2.3.16-2.3.19). So the recursive

MAP estimator algorithm can be used as a different form of the Kalman filter. And

since the KF equations can be derived from recursive MAP estimator, so this work

shows a special way to derive KF algorithm.

Chapter 3

State Estimation of Ballistic Target

Missile from Seeker Measurements
using Nonlinear Filters

This chapter examines the estimation of relative position and relative velocity of a

tactical ballistic missile (target) with respect to a tracking homing missile (inter-

ceptor) from noisy radio frequency (RF) seeker measurements when the engagement

takes place in exo-atmosphere. An extended Kalman filter (EKF), unscented Kalman

filter (UKF) and particle filter (PF) seperately estimate the states of the target. The

performances of the EKF, UKF and PF have been compared. Considering the esti-

mation error and computation load EKF can be considered as a most optimal filter

for estimation state of a ballistic target in exo-atmosphere.

3.1 Introduction

The problem of tracking ballistic missile has attracted much attention in literature

due to both practical and theoretical reasons. The tactical ballistic missile (TBM)

flight can be broadly divided into three phases: 1. boost phase, 2. mid-course

phase (ballistic phase), 3. terminal phase (re-entry Phase). During the boost phase

33
34

the TBM experiences the forward acceleration due to thrust. Moreover the rate of

change of acceleration (jerk) is considerable as the burn time of the booster increases.

In the Ballistic phase the target is well above the atmosphere and it experiences

only gravitational force. During the Re-entry phase it experiences rapid drag decel-

erations. Many tracking filters based on radar measurements have been developed.

The paper [44], compared the error performance of nonlinear filters (EKF, UKF and

PF) for tracking from radar observations of a ballistic re-entry object which has a

one-dimensional vertical motion model with unknown ballistic co-efficient. The per-

formance of the above nonlinear filters was also compared in [45] for the problem of

tracking by processing radar measurements of a 2-dimensional ballistic object with

unknown ballistic co-efficient in the re-entry phase. [46] presented an improved se-

quential Monte-Carlo filter for ballistic target tracking with a random time varying

ballistic co-efficient in the re-entry phase from radar measurements. These problems

have important applications in missile defence system and safety of ageing satellites

and reusable space vehicles during reentry.

This chapter deals with the problem of tracking a tactical ballistic missile (using

the measurements from an RF seeker mounted on a homing ballistic missile) during

the endgame part of interception. A tactical ballistic missile (target) is to be inter-

cepted by a homing ballistic missile (interceptor). The interception takes place during

the terminal phase of flight of the interceptor while the target is in its mid-course

phase, also the interception takes place in the exo-atmospheric region. During this

phase of the flight of the target, the predominant force acting on the target is earth’s

gravitational force, while the effect of atmospheric drag can be ignored. In [47] track-

ing filters are considered for estimation of the line-of-sight (LOS) rate. Paper [48]
 35

describes the modeling of an imaging seeker and an EKF for estimation of LOS rates

using measurements of relative angular displacement between seeker gimbals and a

low-cost strap down inertial unit.

The objective of the present work is to estimate the relative position and relative

velocity of the target using measurements obtained from RF seeker. The measured

signals are contaminated by high degree of noise due to eclipsing, glint, radar cross

section fluctuation and thermal effect [49]. The interceptor missile kinematic informa-

tion is always available from the strap down inertial navigation system (SDINS). The

aim is to formulate nonlinear tracking filters e.g. EKF [13]-[16], UKF [30]-[33], PF

[17] - [20] to estimate the relative kinematic of tactical ballistic missile w.r.t homing

missile using RF seeker measurements and compare the performance of the tracking

filters. The next section describes the relative kinematics of the target and missile (in

Cartesian co ordinates) in the inertial frame for designing the EKF, UKF and PF. The

measurements are obtained in the inner gimbal frame [49, 50] in polar form. Since

the state equations are in Cartesian form while measurement equations are in polar

form, it is a CP formulation. The mathematical modeling of state and measurement

model are given in Sec. 3.2. Sec. 3.3 contains simulation results and conclusions are

drawn in Sec. 3.4.

3.2 Mathematical Modeling

The present objective is to estimate the relative kinematics of the target missile from

noisy measurements obtained from the RF seeker mounted on the interceptor missile.

The signals measured are the range rate, the LOS rates and the gimbal angles.

Here the mathematical modeling of the state estimation problem is presented in

36

the CP - Cartesian -Polar frame. The six states - namely, the relative positions

and velocities along the three axes, are in the inertial frame (cartesian form). The

measurements obtained from the RF seeker are in the polar inner gimbal frame. Strap

down inertial navigation system (SDINS) onboard the interceptor missile gives its

acceleration data in the local level frame, which is considered to be inertial (neglecting

rotation of earth). In this problem, therefore, various data are obtained in different

reference frames. These are the LOS frame, the inner gimbal frame, the navigation

frame, etc. Direction cosine matrices (DCM) are thus required for converting data

from one frame to another frame [49, 50].

Consider Fig. 3.1 The inertial axes system Xi -Yi -Zi are directed along vertically

up, east and north respectively. Let the missile and target position and velocity

components in inertial frame be (xm , ym , zm , Vmx , Vmy , Vmz )T and (xt , yt , zt , Vtx ,

Vty , Vtz )T respectively. So, the relative position and velocity components of missile to

target are

∆x = xt − xm

∆y = yt − ym

∆z = zt − zm

∆Vx = Vxt − Vxm

∆Vy = Vyt − Vym

∆Vz = Vzt − Vzm

Missile acceleration components [amx , amy , amz ]T are obtained from SDINS on-

board interceptor missile and the target acceleration denoted by [atx , aty , atz ]T are

due to the gravitational force only.

 37

The LOS axes system consists of X1 -Y1 -Z1 as shown in Fig 3.1.

Figure 3.1: Axes convention of LOS frame w.r.t inertial frame

LOS angle and LOS rate along azimuth and elevation plane of LOS Frame The

LOS frame is obtained from the inertial frame by means of two successive rotations.

The first is about Xi by an angle −λa and the next is about the newly obtained Y

axis by an angle (900 − λe ) (shown in Fig. 3.2). Where (λa and λe are LOS angle

along azimuth and elevation plane of LOS Frame respectively. The DCM of Inertial
38

Figure 3.2: Rotation sequence used to obtain LOS frame from inertial frame
 39

frame to LOS frame is therefore

 
sin λe cos λe sin λa cos λe cos λa
 
Cl = 
i
 0 cos λa − sin λa 
 (3.2.1)
− cos λe sin λe sin λa sin λe cos λa

In the LOS frame, the range (r), range rate (ṙ), azimuth angle (λa ), elevation an-

gle (λe ), azimuth angle rate (λ˙e ) and elevation angle rate (λ˙a ) respectively, may be

expressed in terms of the relative position and velocity as (Fig. 3.1):

p
r= ∆x2 + ∆y 2 + ∆z 2 (3.2.2)

∆x∆ẋ + ∆y∆ẏ + ∆z∆ż

ṙ = p (3.2.3)
∆x2 + ∆y 2 + ∆z 2

∆y
λa = tan−1 (3.2.4)
∆z
∆x
λe = tan−1 p (3.2.5)
∆y 2 + ∆z 2

(∆z∆ẏ − ∆y∆ż)
λ˙a = (3.2.6)
(∆y 2 + ∆z 2 )

∆ẋ∆y 2 + ∆z 2 − ∆x(∆y∆ẏ + ∆z∆ż)

λ˙e = p (3.2.7)
(∆x2 + ∆y 2 + ∆z 2 )( ∆y 2 + ∆z 2 )
(3.2.2)-(3.2.7) describe the variables in the LOS frame. These are converted in the

inner gimbal frame through the DCM’s. The DCM matrix for converting a vector

from inertial to body frame is expressed through the body quaternion. If the body

quaternion be ~q = q1 i + q2 j +q3 k +q4 , then the DCM from body to inertial frame is:
 
(q42 + q12 − q22 − q32 ) 2(q1 q2 + q3 q4 ) 2(q1 q3 − q2 q4 )
 
Cib = 
 2(q1 q2 − q3 q4 ) (q42 − q12 − q22 − q32 ) 2(q2 q3 + q1 q4 ) 
 (3.2.8)
2(q1 q3 + q2 q4 ) 2(q2 q3 − q1 q4 ) (q42 + q12 − q22 − q32 )
40

where i, j, k, are the unit vectors along xb , yb , zb axes of the body frame respec-

tively. The fin frame conversion is obtained from body frame by 450 rotation about

xb (shown in Fig. 3.3). The DCM of fin frame with respect to body frame is
 
1 0 0
 
Cbf =  1
 0 √2 √1
2

 (3.2.9)
0 − √12 √1
2

The fin dash frame on which the seeker is mounted is shown in Fig 3.4. DCM of fin

dash frame with respect to fin frame is

 
1 0 0
 
Cff d = 
 0 −1 0 
 (3.2.10)
0 0 −1

Finally, the inner-gimbal frame of the seeker is obtained from fin-dash frame by

means of two successive rotations. First rotation is about Yf d (−Yf ) axis by an angle

φy and next rotation is about newly obtained Zf d axis by an angle φz (Fig. 3.4).

The gimbal angles (φy , φz ) in findash frame are shown in Fig 3.4. Here the rota-
φy φz
tion sequence is (xf d yf d zf d )−→(x 1 yg zf d ) −→(xg yg zg ). So the DCM from fin-dash

frame to inner gimbal frame is

 
cos Φy cos Φz sin Φz − cos Φz sin Φy
 
Cfgd = 
 − cos Φy sin Φz cos Φz sin Φy sin Φz 
 (3.2.11)
sin Φy 0 cos Φy

The gimbal angle φy and φz can be calculated as follows:

Considering l, m, n to be unit LOS vectors in the fin dash frame. Then using (3.2.1),

(3.2.8)-(3.2.11) the LOS vector in the LOS frame may be converted to inner gimbal
 41

Figure 3.3: Axes Convention of interceptor body and fin frame

Figure 3.4: Interceptor findash and seeker frame

42

frame as:
       
l 1 1 cos Φy cos Φz
       
 m  = (C g C f d Cfb Cib )  0  = (C g )T  0  =  sin Φz  (3.2.12)
  fd f   fd    
n 0 0 − cos Φz sin Φy
From the above equation the gimbal angles φy and φz can be calculated as:
n m
Φy = tan−1 (− ) & Φz = tan−1 ( √ ) (3.2.13)
l l 2 + n2
Let −
→
ωl be the LOS rates in LOS frame while −
→ be the LOS rates in inner gimbal
ω g

frame. Once more using (3.2.1), (3.2.8)-(3.2.11) as derived above, we get,

   
ωgx −λ˙a sin λe
→=
− 
 ωgy  = [C g C f d C f Cib Cli ] −
→ 
ωl = Cfgd Cff d Cbf Cib Cli 

 (3.2.14)
ωg   fd f b  λ˙e 
ωgz ˙
λa cos λe
where ωgy and ωgz are LOS rate of target along azimuth and elevation of inner gimbal

frame of interceptor. Finally, (3.2.3), (3.2.13) and (3.2.14) give the measurements in

the inner gimbal frame in terms of states defined by (3.2.15) in LOS frame. These

are therefore the measurement equations of the filter.

3.2.1 State Space Equations

The six state variables are the relative position (∆x, ∆y, ∆z), relative velocity (∆Vx , ∆Vy , ∆Vz ))

of the target missile with respect to the tracking missile (interceptor) along X, Y and

Z-axis respectively (Fig 3.1).

      
∆x 0 0 0 1 0 0 ∆x 0 0 0
      
 ∆y   0 0 0 0 1 0   ∆y   0 0 0   
       a −a
       mx tx
d  ∆z 
 
= 0 0 0 0 0 1 

 ∆z  
+ 0 0 0   
  amy − aty +w(t)
dt   
 ∆Vx   0 0 0 0

0 0 
 
∆Vx   −1 0

0 

       amz − atz
 ∆V   0 0 0 0 0 0   ∆Vy   0 −1 0 
 y      
∆Vz 0 0 0 0 0 0 ∆Vz 0 0 −1
(3.2.15)
 43

where, w(t) is assumed Gaussian zero mean white process noise. (amx , amy , amz ) are

interceptor acceleration components in inertial frame obtained from missile inertial

navigation system (INS) data. The target acceleration is due to only the gravitational

force, along X, Y and Z-axis and is denoted by (atx , aty , atz ) respectively.

3.2.2 Measurement Equations

Measurements are defined in inner gimbal frame. Using (3.2.3), (3.2.13) and (3.2.14),

and adding a suffix ‘m0 to denote measurements of seeker, we have,

∆x∆ẋ + ∆y∆ẏ + ∆z∆ż

ṙm = p + η2 (3.2.16)
∆x2 + ∆y 2 + ∆z 2
n
Φym = tan−1 (− ) + η3 (3.2.17)
l
m
Φzm = tan−1 ( √ ) + η4 (3.2.18)
l2 + n2
 
   
−   −
   −λ sin
˙ λe   
 ωgym  = C g C f d C f Cib Cli 

a  
 +  η5  (3.2.19)
  f d f b 

 λ˙e 

ωgzm η6
˙ λe
λa cos
where , η2 , η3 , η4 , η5 & η6 are measurement noise, assumed to be zero mean white

Gaussian and uncorrelated with process noise.

It may be noted that the state equation is linear but the measurement equations

are nonlinear.

3.3 Simulation Results

The relative position and velocity of target missile with respect to interceptor missile

have been estimated using state space equation (3.2.15) and measurement equations
44

(3.2.16-3.2.19) given in Section 3.2. EKF, UKF and PF are used as tracking filters

and the performance of these filters have been compared from the simulation results.

Sampling time of the seeker measurement is 25 ms. Number of particles is chosen

to be 1000 for the PF. The initial state covariance, process noise covariance and the

measurement noise covariance for all the filters are same and selected as follows.

The initial state covariance of relative position, and relative velocity along x-axis,
2 2 2
y-axis and z-axis has been selected respectively as σ∆x = 100, σ∆y = 100, σ∆z = 100,
2 2 2 2
σ∆vx
= 100, σ∆vy
= 100, σ∆vz
= 100. The process noise covariance : σ∆x = 10−1 ,
2
σ∆y = 10−1 , σ∆z
2
= 10−1 , σ∆v
2
x
2
= 1, σ∆vy
2
= 1, σ∆vz
= 1. The measurement noise

covariance has been selected for ṙm , Φym , Φzm , ωgym and ωgzm respectively as σṙ2 = 202 ,
2.5π 2.5π 3π 3π
σφ2 ym = 180
, σφ2 zm = 180
, σω2 gym = 180
, σω2 gzm = 180
The seeker measurements noise of

range rate (ṙ), azimuth angle (λa ), elevation angle (λe ), azimuth angle rate (λ˙e ) and

elevation angle rate (λ˙a ) in inner gimbal frame are shown in Figs. 3.5- 3.9.

noise in Range rate

10
Measure − Actual

0
metre/sec

−5

−10

−15

−20

−25
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 3.5: Noise in range rate

 45

noise in gimbal angle along azimuth

0.5

0
degree

−0.5

−1

−1.5

Measure − Actual
−2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
normalized time

Figure 3.6: Noise in gimbal angle along y-axis

noise in gimbal angle along elevation

1.5
Measure − Actual

0.5
degree

−0.5

−1

−1.5
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 3.7: Noise in gimbal angle along z-axis

46

noise in LOS rate along azimuth

Measure − Actual
8

4
degree/sec

−2

−4

−6

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

normalized time

Figure 3.8: Noise in LOS rate along y-axis

noise in LOS rate along elevation

8 Measure − Actual

2
degree/sec

−2

−4

−6

−8

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

time in sec

Figure 3.9: Noise in LOS rate along y-axis

 47

estimation error in position along X−axis

25
EKF
UKF
PF
20

15
metre

10

0
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 3.10: Comparison of estimation error of relative position along X-axis using
EKF, UKF and PF

estimation error in position along Y−axis

20

15

10

0
metre

−5

−10

−15

−20
EKF
−25 UKF
PF
−30
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 3.11: Comparison of estimation error of relative position along Y-axis using
EKF, UKF and PF
48

estimation error in position along Z−axis

10

−5
metre

−10

−15

−20 EKF
UKF
PF
−25
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 3.12: Comparison of estimation error of relative position along Z-axis using
EKF, UKF and PF

estimation error in velocity along X−axis

5

−5
metre/sec

−10

−15
EKF
UKF
PF
−20
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 3.13: Comparison of estimation error of relative velocity along X-axis using
EKF, UKF and PF
 49

estimation error in velocity along Y−axis

8
EKF
6 UKF
PF
4

2
metre/sec

−2

−4

−6

−8

−10
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 3.14: Comparison of estimation error of relative velocity along Y-axis using
EKF, UKF and PF

estimation error in velocity along Z−axis

8

2
metre/sec

−2

−4

−6
EKF
−8 UKF
PF
−10
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 3.15: Comparison of estimation error of relative velocity along Z-axis using
EKF, UKF and PF
50

3.4 Conclusion

From the plots of estimation error (Fig. 3.10-Fig. 3.15) we can see that results of the

filters are comparable. Though in some cases (relative position and relative velocity

along y-axis) the results using PF is better from the accuracy point of view. But

the computational load of PF is much higher than UKF and EKF. The approximate

relation of computational load is given in the first chapter in Table 1.1. The estimation

error using UKF and EKF is almost same as in this case the state space equation of

the system is linear though the measurement equations are nonlinear but the time

step is quite small i.e. 25 ms. So the error due to linearization is less for this system

model compared to the system model shown in Chapter 1 , Section 1.3.1 where both

the state space and observation model were nonlinear with a time step of 1 sec. But

considering both the computational burden and accuracy of the state estimation, we

can say than the EKF is the best choice for state estimation of ballistic target in

exo-atmospheric region using RF seeker measurements.

Chapter 4

Seeker Filter Design of Ballistic

Target Missile Incorporating
Seeker Dynamics

This chapter examines a method to estimate states using the sensor measurements

from an RF seeker which contains lag due the mechanism of data collection and

transmission to the estimator in the system. The error due to the measurement lag

can be handled by incorporating the sensor dynamics in the system model. In this

chapter, an extended Kalman filter has been used to estimate relative kinematics of

a ballistic target missile in the exo-atmospheric region w.r.t. interceptor missile from

noisy and lagged seeker measurements. The seeker dynamics has been modeled and

included in the system model of the estimator in order to handle the lag present in the

seeker measurements due to processing and transmitting data obtained in different

reference frames.

4.1 Introduction

In target tracking problems often measurements are obtained from various sensors in

different reference frames. Usually during processing and transmitting of data, time

51
52

lag is introduced in them before the same reaches the estimator (filter). In the filtering

equations, one thus faces problems in updating the states in current time step with

measurements containing time lag. Several algorithms have been developed to cope

with this difficulty. The initial work of [51] and [52] presented an approximate solution

to the out of sequence measurement problem, to be further referred to as ”algorithm

B”. The optimal solution to this problem, referred to as ”algorithm A”, was derived

in [53]. The works of [54] and [55] solved the updating problem of out of sequence

measurements with an arbitrary lag. Yet they both of them required a backward

iteration for l step and also considerable amount of storage. In the present work a

ballistic target is intercepted in the exo-atmospheric region as has been mentioned in

previous chapter (Sec. 3.2). The technique developed here to handle time lag in the

measurement does not require data storage. The main idea is to model and include

the entire sensor mechanism, i.e. the seeker dynamics, with its track and stabilization

loop [56, 58] (named seeker model), in the estimator design. This compensates for

the lag and error present in the measurements caused by the seeker dynamics. In

paper [57], unlike a conventional Kalman filter (where the target kinematics only is

assumed as a plant description), the seeker dynamics with its track and stabilization

loop are included in the plant and then the platform motion is taken to be a known

input.

In the present work, the conventional extended Kalman filter [13]-[16] has been

used as the estimator. In the plant model the seeker states have been augmented

with the target states. Inclusion of the simplified seeker dynamics in estimator design

reduces the estimation error produced by the lag and uncertainty arising due to the

seeker dynamics. From simulation results it may be seen that inclusion of seeker
 53

dynamics in the plant model improves the estimator performance.

4.2 Problem Formulation

System model:

A general nonlinear system and observation model has been considered

ẋ(t) = f (x, t) + w(t) (4.2.1)

y(t) = h(x(t), t) + v(t) (4.2.2)

where x(t) are system states to be estimated from noisy measurements, y(t), in turn

obtained from the sensor, and w(t), v(t) are zero mean white Gaussian noise.

The sensor (generating y(t)) has its own dynamics, not shown here. During esti-

mator design, sensor dynamics is generally neglected. In case it is possible to incor-

porate the sensor dynamics suitably in the state model, it may compensate for the

extra error caused by the presence of time lag in the measurement. And the estimate

will be more accurate. To examine this, a system is considered below which contains

a sensor with a general 2nd order dynamics. It is shown how the effect of sensor

dynamics may be included in the the system model during estimator design.

In the continuous time domain, the measurements entering the estimator are ym (t)

and noise present in the measurement data are assumed to be v(t). Without taking

the sensor dynamics into account we may write,

ym (t) = yt (t) + v(t) (4.2.3)

where yt (t) is the true measurement. Next considering 2nd order sensor dynamics (Fig

4.1), the actual measurement data (ya (t)) coming out of the sensor can be expressed

in frequency domain (s-domain) as

54

Figure 4.1: 2 nd order transfer function

ωn2
Ya (s) = Yt (s) (4.2.4)
s2 + 2ξωn s + ωn2
s2 Ya (s) + 2ξωn sYa (s) + ωn2 Ya (s) = ωn2 Yt (s) (4.2.5)

Taking inverse Laplace transform

ÿa (t) + 2ξωn ẏa (t) + ωn2 ya (t) = ωn2 yt (t) (4.2.6)

To incorporate the sensor dynamics in the system model ((4.2.1), (4.2.2)), the sen-

sor states (4.2.6) are augmented with the original state model (4.2.1). The system

incorporating sensor dynamics, i.e., the augmented state and observation model can

be represented as follows.

Augmented state model:

   
x(t) f (x(t), t)
d   
 ya (t)  = 

 + w(t)
   y˙t (t)  (4.2.7)
dt
2 2
y˙a (t) −2ξωn ẏa (t) − ωn ya (t) + ωn yt (t)

Where yt (t) = h(x(t), t) , from (4.2.2)

And observation model is

ym (t) = ya (t) + v(t) (4.2.8)

 55

It may be noted that ym (t) are the measurements entering the sensor and in (4.2.8),

these effect of time lag is included. This is a more accurate representation compared

to (4.2.3). The estimator designed using system model (4.2.7)-(4.2.8) can be used to

handle the error due to measurement lag caused by sensor dynamics.

4.3 Tracking Filter Design incorporating Seeker

Model

The problem described in the previous section is illustrated using a ballistic missile.

The missile tracks its target using a range and angle sensing system known as the

seeker. This device is mounted on an instrumentation platform residing in the forward

part of the missile. Seeker measurements are used to estimate the states of the ballistic

target missile. The state and observation models of the ballistic target as well as the

interception model are given in the previous chapter (Sec. 3.2).

Modeling of Seeker Dynamics

The schematic model of the seeker (from [56]) is shown in Fig 4.2. Input to the seeker

block is the true kinematic LOS rates in inner gimbal frame (ωgy , ωgz ). These are

processed through the seeker loop of Fig 4.2 [56], to generate the measured LOS rate

in the inner gimbal frame (ωgym , ωgzm ). These measurements get contaminated by

measurement noise. The resultant noise is produced due to glint, radar cross section

(RCS) fluctuation and thermal effect. Also due to eclipsing effect in seeker, which is a

function of pulse repetitive frequency (PRF), LOS rates are not available periodically.

These effects are represented in the simulation model shown in Fig 4.2. True angle

tracking errors (εy , εz ) are contaminated by receiver noise and glint noise and are

passed to the signal processing block (GDSP (s)). To include the functionality of the
56

Figure 4.2: Block diagram of seeker model

DSP block, GDSP (s) is assumed to be a mean value of 9 pulses within 10 ms period.

From the block diagram of the seeker model we can write.

GDSP (s)Kv
ωgm (s) = (ωgt (s) − Gs ωgm (s)) + v(s)GDSP (s)Kv
s
(ωgt (s) − Gs ωgm (s))
= [ + v(s)]GDSP (s)Kv (4.3.1)
s
ω1
In this seeker model, typical value of ω1 of the stabilization loop (TF Gs (s) = s+ω1
)

is 100 rad/sec, and hence its effect in estimation is neglected. i.e. Gs (s) ≈ 1

By system identification it is seen that the DSP block can be represented by the

following transfer function:

KDSP
GDSP (s) = 1+TDSP s
So (4.3.1) can be further simplified as

(ωgt (s) − ωgm (s))

ωgm (s) = [ + v(s)]GDSP (s)KV
s
KDSP KV × ωgt (s) sKDSP KV × v(s)
ωgm (s) = + (4.3.2)
TDSP s2 + s + KDSP KV TDSP s2 + s + KDSP KV
 57

To study the effect of modeling of seeker dynamics on the estimates, the effect of

seeker dynamics on the measurement noise has been neglected here.

sKDSP KV ×v(s)
Hence TDSP s2 +s+KDSP KV
≈ v(s)

LOS rate measurements are normally (without considering the seeker dynamics) given

by

LOS measurement = True LOS rate + noise

From (4.3.2) we can see that due to the seeker dynamics the actual value of the LOS

rate (ωga ) may be written as

KDSP KV × ωgt (s)

ωga (s) =
TDSP s2 + s + KDSP KV
ω 2 × ωgt (s)
= 2 n (4.3.3)
s + 2ξωn s + ωn2
q
KDSP KV
where natural frequency of oscillation of the seeker model is ωn = TDSP
and
1
damping ratio ξ = √
2 KDSP KV TDSP

In time domain (4.3.3) can be represented as

ω̈ga (t) + 2ξωn ω̇ga (t) + ωn2 ωga (t) = ωn2 ωgt (t) (4.3.4)

System Model Incorporating Seeker Dynamics

The state model (3.2.15) is augmented with the additional seeker states, namely,

the actual LOS rate and its higher order derivatives (4.3.4) (ω̈ga (t), ω̇ga (t), ωga (t))

to incorporate seeker dynamics in the state model. The six base state variables

are relative position (∆x, ∆y, ∆z) and relative velocity (∆Vx , ∆Vy , ∆Vz ). The seeker

dynamics includes six augmented states comprising of the two LOS rates and their
58

higher order derivatives along y-axis and z-axis.

     
∆x ∆Vx 0 0 0
     
 ∆y   ∆V   0 0 0 
   y   
     
 ∆z   ∆V   0 0 0 
   z   
     
 ∆Vx   0   −1 0 0 
     
     
 ∆Vy   0   0 − 1 0  
      amx − atx
d   
 ∆Vz   0  
  0 0−1   
 = +  a
 my − a 
ty +w(t)
dt  ωgya   ω̇   0 0 0  
   gya  
      amz − atz
 ω̇gya   −2ξωn ω̇gya − ωn2 ωgya + ωn2 ωgyt   0 0 0 
     
     
 ω̈gya   0   0 0 0 
     
 ω   ω̇gza   0 0 0 
 gza     
     
 ω̇gza   −2ξωn ω̇gza − ω ωgza + ω ωgzt  
2 2
0 0 0 
   n n   
ω̈gza 0 0 0 0
(4.3.5)

q
KDSP KV 1
ωn = TDSP
is natural frequency of oscillation and ξ = √
2 KDSP KV TDSP
is damping

ratio of the seeker model, where KDSP and TDSP are gain and time constant of the

DSP block respectively. The track loop gain, KV , is time varying, evaluated online

from simulation environment. The true LOS rates (ωgyt , ωgzt ) in the inner gimbal

frame are computed as given below.

     
− −λ˙a sin λe −
     
 ωgyt  = C g C f d C f Cib Cli  λ˙e  +  η5  (4.3.6)
  fd f b    
ωgzt λ˙a cos λe η6

Where,
∆y
λa = tan−1 (4.3.7)
∆z
∆x
λe = tan−1 p (4.3.8)
∆y 2 + ∆z 2
 59

(∆z∆Vy − ∆y∆Vz )
λ˙a = (4.3.9)
(∆y 2 + ∆z 2 )
∆Vx (∆y 2 + ∆z 2 ) − ∆x(∆y∆Vy + ∆z∆Vz )
λ˙e = p (4.3.10)
(∆x2 + ∆y 2 + ∆z 2 )( ∆y 2 + ∆z 2 )
The observation model in inner gimbal frame is:
     
∆x∆Vx +∆y∆Vy +∆z∆Vz
ṙ √ η
 m   ∆x2 +∆y 2 +∆z 2   1 
 Φ     η 
 ym   tan−1 (− nl )   2 
     
 Φzm  =   +  η3  (4.3.11)
   tan−1 ( √l2m+n2 )   
     
 ωgym   ωgya   η4 
     
ωgzm ωgza η5

4.4 Simulation Results

The states of the target missile are estimated first using the system including the

seeker model ((4.3.5)-(4.3.11)) and next it is compared with the estimation results

based on the system without the seeker model ((3.2.15)-(3.2.19)). Sampling time of

the seeker measurement is 25 ms. The initial state covariance, process noise covariance

and the measurement noise covariance for the system without seeker model have

been selected as follows. The initial state covariance of relative position and relative
2
velocity along x-axis, y-axis and z-axis have been selected respectively as σ∆x = 100,
2 2 2 2 2
σ∆y = 100, σ∆z = 100, σ∆vx
= 100, σ∆vy
= 100, σ∆vz
= 100.
2
The process noise covariance : σ∆x = 10−4 , σ∆y
2
= 10−4 , σ∆z
2
= 10−4 , σ∆v
2
x
= 10−1 ,
2
σ∆vy
= 10−1 , σ∆v
2
z
= 10−1 .

The measurement noise covariance for both the systems has been selected for ṙm , Φym ,
2.5π 2.5π 3π
Φzm , ωgym and ωgzm respectively as σṙ2 = 202 , σφ2 ym = 180
, σφ2 zm = 180
, σω2 gym = 180
,
3π
σω2 gzm = 180
.

In case of the target model with augmented state of seeker model, the initial values
60

are selected as: The initial state covariance of relative position, relative velocity along

x-axis, y-axis and z-axis, LOS rates, higher order derivatives of LOS rates along y-
2 2 2
axis and z-axis has been selected respectively as σ∆x = 100, σ∆y = 100, σ∆z = 100,
2 2 2
σ∆vx
= 100, σ∆v y
= 100, σ∆vz
= 100, σωgya = 10−1 , σω̇gya = 10−1 , σω̈gya = 10−1 ,

σωgza = 10−1 , σω̇gza = 10−1 , σω̈gza = 10−1 .

2
The process noise covariance : σ∆x = 10−4 , σ∆y
2
= 10−4 , σ∆z
2
= 10−4 , σ∆v
2
x
= 10−1 ,
2
σ∆vy
= 10−1 , σ∆v
2
z
= 10−1 , σωgya = 10−4 , σω̇gya = 10−4 , σω̈gya = 10−1 , σωgza = 10−4 ,

σω̇gza = 10−4 , σω̈gza = 10−1 .

q
KDSP KV
Natural frequency of oscillation of the seeker model is ωn = TDSP
and damping
1
ratio ξ = √
2 KDSP KV TDSP
, where KV is a time varying gain, evaluated online from

simulation environment. Gain of the DSP block KDSP = 1 and TDSP = 0.0125

The plots of absolute values of estimation error (estimated state - true state) of

relative position and relative velocity along x, y and z-axis (Figs. 4.3-4.8 respectively)

are given for both the models with and without seeker dynamics.
 61

Figure 4.3: Comparison of absolute value of estimation error of relative position along
X-axis with and without incorporating seeker model

Figure 4.4: Comparison of absolute value of estimation error of relative position along
Y-axis with and without incorporating seeker model
62

Figure 4.5: Comparison of absolute value of estimation error of relative position along
Z-axis with and without incorporating seeker model

Figure 4.6: Comparison of absolute value of estimation error of relative velocity along
X-axis with and without incorporating seeker model
 63

Figure 4.7: Comparison of absolute value of estimation error of relative velocity along
Y-axis with and without incorporating seeker model

Figure 4.8: Comparison of absolute value of estimation error of relative velocity along
Z-axis with and without incorporating seeker model
64

4.5 Conclusion

In this chapter a seeker model consisting of seeker dynamics with its track and stabil-

ity loop is used in a ballistic target interception problem. A state-augmented seeker

model is integrated with the original system. A seeker filter is then designed using

an EKF for this state augmented system. The inclusion of the seeker model com-

pensates the error due to time lag in the measurement caused by the processing and

transmitting process. The simulation results compare the estimation performances

(in terms of estimation error) between the models with and without seeker dynamics.

From the plots it can be seen that during the initial time span the estimation error is

comparable in both the cases but during final time span there seems to be improve-

ment in the performance of the estimator, working with the seeker model, compared

to the one without the seeker model. This may be because of the fact that during

the final time the LOS rate is changing fast and hence the seeker loop dynamics is

predominant during this time span. The strategy used here may also be applied to

similar problems where there is considerable time lag in the measurement due to the

sensor dynamics.
Chapter 5

Kinematic State Estimation of a

Ballistic Target Missile from Bad
Seeker Measurements

The problem is to estimate realtime the kinematic states of a ballistic missile from

seeker measurements containing jumps, discontinuity and spiky noise i.e. bad mea-

surements. A modified extended Kalman filter (EKF) detects the instant when a bad

measurement is received and tunes the value of measurement noise covariance (R)

accordingly. This modified EKF with a time varying noise covariance gives better re-

sults compared to conventional EKF. The seeker measurements have been generated

in a 6-degree of freedom (DOF) simulation environment.

5.1 Introduction

The objective here is realtime state estimation of a ballistic missile with respect to a

tracking missile using seeker measurements containing a high degree of uncertainty.

The seeker measurements generated in a 6 DOF simulation program, represent signals

which are contaminated by high degree of noise due to eclipsing, glint, radar cross

section fluctuation and thermal effect [49]. These noise appear in the form of jumps,

65
66

spikes, discontinuity and data loss over and above the usual zero mean white Gaussian

noise present in a radio frequency (RF) seeker.

Some work has been reported in the literature to handle missing measurements

(i.e. data loss or discontinuity in the measurements) and spiky noise in measurements.

In [59], a filtering problem for linear uncertain discrete-time stochastic systems with

missing measurements was considered with norm bounded parameter uncertainties in

the state matrix and the probability of occurrence of missing data is assumed to be

known. In the papers [60] - [61], a robust filtering problem was studied for stochastic

uncertain discrete time-varying systems with missing measurements described by a

binary switching sequence which in turn satisfied a conditional probability distribu-

tion. For navigation and tracking applications, the paper [62], considered the problem

of Kalman filtering with intermittent observations by modeling the instant of arrival

of the observation as a random process.

A spiky noise appearing in radar measurements of the target tracking system [63]

can be characterized by glint noise. This glint noise has a long-tailed distribution [64]

- [65] and can severely degrade the performance of a Kalman filter. In 1975, Masreliez

[66]- [67] introduced a score function based scheme to deal with glint noise. While

this approach looked promising, he encountered some implementation problems. In

[63], [68], the author handled glint noise via a score function approximation and

applied it to target tracking problems. He also incorporated the Masreliez filter

into the interacting multiple model (IMM) algorithm and obtained a nonlinear IMM

algorithm [69]. The paper [70], characterized glint noise as a mixture of two Gaussian

components and used two different models to represent the noise arising from these

two Gaussian components. Hence among these several methods existing to handle
 67

different types of uncertainty in the measurements, some of them are computationally

expensive. In this work a simple strategy has been implemented in the KF algorithm

to handle all types of uncertainties present in the measurements.

Spiky measurements with jumps, discontinuity and data loss are termed ‘bad

measurements’. The model formulation has been done in CP frame, discussed in

Sec 3.2. It has been observed that in each case the estimated state deviates from the

true value due to jump or spike or data loss presents in the measurements. Here a new

strategy is proposed to handle the situation when the measurement is bad and this

strategy is implemented in EKF. An adjustable noise covariance has been employed

- one is of higher value and another is of lower value. On detection of jump or spike

or data loss in the measurement the noise covariances changes to higher value and it

implies that the system has low confidence in measurement, i.e. the system disregards

those measurements. This strategy has been implemented in the extended Kalman

filter and very good performance of this modified EKF has been obtained compared

to EKF.

5.2 Proposed strategy to handle bad measurement

The measurement noise covariance, R, can be adjusted on detection of jump, discon-

tinuity or spike in the noise. If R is very large it means we disregard the current

measurement while updating the new estimate. This is as expected, i.e., if the mea-

surement noise covariance is large, we have low confidence in the measurement and

our estimate will depend more upon the previous samples of estimates. It has been

implemented in EKF. This is described below:

We know,
68

Innovation = (measurement − predicted measurement),

ν = Zm − Ẑ

A threshold value C of this innovation is selected. The value of C is selected by

observing the pattern of measurements noise history. The jump or spike is detected

in the measurement if ν > C. Two different values of R is selected one is low Rlow

and other is high value Rhigh . If innovation is less than the chosen threshold i.e., if

ν < C, then lower value of noise covariance is selected: R = Rlow .

Otherwise higher value of R is selected i.e., R = eRlow , where, e > 1. This approach

has been implemented in EKF.

5.3 The EKF Algorithm implementing New Strat-

egy

This section contains the modified EKF algorithm to handle bad measurement data,

i.e. when the uncertainty in measurement is very high.

• A priori mean

x̂k|k−1 = φk−1 x̂k−1|k−1 (5.3.1)

• a priori covariance

Pk|k−1 = φk−1 Pk−1|k−1 φk−1 + Qk−1 (5.3.2)

• the innovation

ν = (zk − Hk x̂k|k−1 ) (5.3.3)

• If ν < C

R = Rlow
 69

else

R = eRlow

where e > 1

• a posteriori mean

x̂k|k = x̂k|k−1 + Kk [zk − Hk x̂k|k−1 ] (5.3.4)

• a posteriori covariance

Pk|k = [I − Kk Hk ]Pk|k−1 (5.3.5)

• gain matrix

Kk = Pk|k−1 HkT (Hk Pk|k−1 HkT + Rk )−1 (5.3.6)

5.4 Simulation Results

The relative position and velocity of target missile with respect to tracking is esti-

mated using the system model and measurement model given in Sec 3.2. Sampling

time of the seeker measurement is 25 ms. The state has been estimated using EKF

and the EKF algorithm employing new strategy. The initial state covariance, pro-

cess noise covariance and the measurement noise covariance in all the cases has been

selected as follows.

For both the estimators, the initial state covariance of relative position and relative
2
velocity along X-axis, Y-axis and Z-axis has been selected respectively as σ∆x = 100,
2 2 2 2 2
σ∆y = 100, σ∆z = 100, σ∆vx
= 100, σ∆vy
= 100, σ∆vz
= 100.
2
The process noise covariance for EKF and new EKF has been selected as σ∆x = 0.1,
2 2 2 2 2
σ∆y = 0.1, σ∆z = 0.1, σ∆vx
= 1, σ∆vy
= 1, σ∆v z
= 1.

The measurement noise covariance for EKF has been selected for ṙm , Φym , Φzm , ωgym
70

2.5π 2.5π 3π 3π
and ωgzm respectively as σṙ2 = 202 , σφ2 ym = 180
, σφ2 zm = 180
, σω2 gym = 180
, σω2 gzm = 180

In case of modified EKF, the above values of measurement noise covariance has been

selected for Rlow . And Rhigh has been taken as Rhigh = 100Rlow . The threshold values

has been selected for ṙm , Φym , Φzm , ωgym and ωgzm respectively as −5 ≥ Cṙ ≥ 5,

−.01 ≥ Cφym ≥ .01, −.01 ≥ Cφzm ≥ .01, −.1 ≥ Cωgym ≥ .1, −.1 ≥ Cωgzm ≥ .1.

The plots of estimation error (estimated state - true state) of relative position and

relative velocity along X, Y and Z-axis (Fig. 5.6-Fig. 5.11 respectively) is given using

EKF and the modified EKF. The plots of the corresponding noise in measurements

i.e. ṙm , Φym , Φzm , ωgym and ωgzm are shown in Figs. 5.1-5.5 respectively.

Figure 5.1: Noise in range rate

 71

Figure 5.2: Noise in gimbal angle along y-axis

Figure 5.3: Noise in gimbal angle along z-axis

72

Figure 5.4: Noise in LOS rate along y-axis

Figure 5.5: Noise in LOS rate along z-axis

 73

estimation error in position along X−axis

30

25

20
metre

EKF
15
modified EKF

10

0
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 5.6: Estimation error in ∆x using EKF and modified EKF

estimation error in position along Y−axis

10

−10

−20
metre

−30

−40

−50

−60
EKF
modified EKF
−70
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 5.7: Estimation error in ∆y using EKF and modified EKF)

74

estimation error in position along Z−axis

25
EKF
20 modified EKF

15

10

5
metre

−5

−10

−15

−20

−25
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 5.8: Estimation error in ∆z using EKF and modified EKF

estimation error in velocity along X−axis

16
EKF
14 modified EKF

12

10

8
metre/sec

−2

−4
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 5.9: Estimation error in ∆vx using EKF and modified EKF
 75

estimation error in velocity along Y−axis

10

0
metre/sec

−5

−10

EKF
modified EKF
−15

−20
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 5.10: Estimation error in ∆vy using EKF and modified EKF

estimation error in velocity along Z−axis

8

4
metre/sec

−2

−4
EKF
modified EKF
−6
0 0.2 0.4 0.6 0.8 1
normalized time

Figure 5.11: Estimation error in ∆vz using EKF and modified EKF
76

5.5 Conclusion

Comparing the plots of estimation error it can be seen that the estimation error in case

of newly implemented strategy in EKF is much less than that of EKF. The estimated

states divert from the true value due to the jump present in the measurement data.

But the modified EKF is gives very good estimation results compared to conventional

EKF. So the modified EKF algorithm can be used for realtime state estimation of

the nonlinear systems in presence of bad measurement i.e. high uncertainty in the

measurement.
Chapter 6

Estimating the Probability Density

Function of a Non-Stationary
Non-Gaussian Noise

The problem of estimating the probability density function (PDF) of a non-stationary

non-Gaussian noise (NSNGN) is addressed. The non-Gaussian noise is modeled using

Gaussian mixture PDFs and an algorithm is proposed to estimate the parameters by

maximizing the log likelihood function. Three simulation results illustrate the validity

and utility of the proposed algorithm for stationary or non-stationary, Gaussian or

non-Gaussian, zero mean or non-zero mean and uni-modal or multi-modal distributed

noise.

6.1 Introduction

In many real world situations a signal transmitted by a sensor contains noise [71]-[73].

At times, this noise is purely random in nature and is represented by its probability

density function (PDF). The performance of an estimator [74] is expected to improve

if it contains appropriate information of the noise. Normally [75] the parameters

of a time varying noisy signal are estimated assuming the noise to be a zero mean

77
78

white Gaussian sequence. However in certain cases like, ultrasonic RF signals [76],

images [77], ECG signals [78], radar clutter [79], radar data [63] etc, the PDF of the

noise sequence is actually non-Gaussian. There have been attempts to model this

non-Gaussian PDF in different ways. In [80], the parameters of non-Gaussian signals

were estimated using non-Gaussian auto regressive (AR) processes based on higher

order statistics. Non-Gaussian PDFs were closely approximated by Gaussian mix-

ture models in [81]-[85]. A noise sequence was modeled in [86] as a Gaussian mixture

process and an expectation maximization (EM) algorithm was employed for its pa-

rameter estimation. The parameters of a non-Gaussian AR model were estimated in

[87] by maximum-likelihood estimation algorithm. In [88] the driving noise sequence

was modeled as a Gaussian mixture process and a general dynamic programming

technique was used for its parameter estimation. In [89], the authors presented a re-

cursive algorithm for identifying the order and parameters of an ARMA model driven

by a non-Gaussian random sequence. In [90] and [91], the work proposed iterative

methods using maximum likelihood for parameter estimation of non Gaussian sig-

nals. The authors, in [92], established the probabilistic and geometric meaning of the

correntropy function that can be useful in non-Gaussian signal processing, especially

for impulsive noise. The paper [93] examined a generalized least absolute deviation

method for parameter estimation of AR signals in a non-Gaussian noise environment.

But, till date all the approaches found in the literature consider non-Gaussian but

stationary noise sequences where the parameters, i.e., means, variances and weight-

ing co-efficients, that define the Gaussian-mixture PDFs, are assumed to be fixed and

these methods are not applicable in cases of non-stationary noise.

 79

In the present work a generalized algorithm is developed to estimate the param-

eters of the PDF of noise which may be stationary or non-stationary, Gaussian or

non-Gaussian, zero mean or non-zero mean, uni-modal or multi-modal. This pro-

posed algorithm may be suitably included in an estimator to improve the efficiency of

estimation. In an earlier attempt [94], parameters of a non-stationary non-Gaussian

noise (NSNGN) were estimated for zero mean noise only. The present work is capable

of dealing with NSNGN sequence which may be zero or non-zero mean. A Gaussian

mixture model is used to model the non-Gaussian noise and the parameters are esti-

mated over time in two stages - a newly proposed a priori update and a posteriori

update. The a posteriori parameters are evaluated by maximization of the log like-

lihood function. The proposed algorithm has been used to estimate the PDF of a

non-stationary Rayleigh distributed noise and a non-stationary glint noise which is

observed in radar tracking systems [95]-[96]. The proposed algorithm has been val-

idated and its utility has also been demonstrated by comparing it with an existing

EM algorithm for bi-modal distributed stationary data. The rest of the chapter is

organized as follows. In Section 6.2 the theory for parameter estimation of NSNGN is

derived. The computational algorithm is summarized in Section 6.3. The simulation

results are given in Section 6.4, and conclusions are made in Section 6.5.

6.2 Parameter Estimation

Let us consider a non-stationary non-Gaussian noise (NSNGN) sequence vk (with

zero or non-zero mean). The parameters of vk are required to be estimated from

n number of individual data history. Here, this NSNGN is modeled by Gaussian

mixture densities and the parameters of noise at each time instant are estimated by
80

maximizing the logarithm of the likelihood function of the data set of n independent

noise sequences.

Let f (vk ), the PDF of vk at time instant k, be characterized by a Gaussian mixture

density [90] of size m given by,

m
X
f (vk ) = pi (k)fi (vk ) (6.2.1)
i=0
Xm
pi (k) {vk − µi (k)}2
= exp[− ] (6.2.2)
i=0
(2π)1/2 σi (k) σi2 (k)

where fi (vk ) is the probability density of vk corresponding to i-th Gaussian component

density in the mixture and µi (k) = [µ1 (k), ...., µm (k)], σi (k) = [σ1 (k), ...., σm (k)],

pi (k) = [p1 (k), ...., pm (k)] are respectively the means, standard deviations and weights

of the individual Gaussian component densities. Also pi ≥ 0, and

m
X
pi (k) = 1. (6.2.3)
i=0

The i-th Gaussian density component of vk , normally distributed with mean µi (k)

and variance σi2 (k), is denoted by N{µi (k), σi (k)} at the k-th time step and is related

to the event Ai (k), in the following way,

Ai (k) = [vk ∼ N{µi (k), σi (k)}] (6.2.4)

with Ai (k), (i = 1, .., m), mutually exclusive and exhaustive, and the probability of

the event Ai (k) being,

P {Ai (k)} = pi (k). (6.2.5)

Using Bayes’ rule, the posterior probability of the i-th Gaussian component condi-

tioned on signal vkl is

p{v l |fi (v l )}P {fi (vkl )}

P {fi (vkl )|vkl } = Pn k l k l l
j=1 p{vk |fi (vk )}P {fi (vk )}
 81

where vkl denotes noise in the l-th data set (among a total of n data sets) at the k-th

time instant and p{vkl |fi (vkl )} is the conditional PDF of vkl given fi (vkl ).

From (6.2.4) we can say, the i-th Gaussian density component of vkl (denoted by

fi (vkl )) is related to the event Ai (k). Now substituting Ai (k) with fi (vkl ) in (6.2.5),

we get

P {Ai (k)} = P {fi (vkl )} = pi (k) (6.2.6)

And, since,

p{vkl |fi (vkl )} = fi {vkl |µi (k), σi (k)}, (6.2.7)

we have
pi (k)fi {vkl |µi (k), σi (k)}
P {fi (vkl )|vkl } = Pm l
. (6.2.8)
j=1 pi (k)fi {vk |µi (k), σi (k)}
Also
m
X
P {fi (vkl )|vkl } = 1. (6.2.9)
i=1
Now the parameters, i.e., pi (k), µi (k) and σi (k), will be estimated by maximizing the

complete data log likelihood w.r.t. the function variables pi (k), µi (k) and σi (k), using

the a priori values of the parameters (a priori update of parameters is discussed at

the end part of this section through (6.2.21)-(6.2.26)). Log likelihood is used instead

of true likelihood because it leads to simpler formulae, but still attains its maximum

at the same points as the likelihood. The logarithm of the likelihood function of n

independent signals conditioned on the model parameters is given by [90]

n
X m
X
Ω(k) = loge pi (k)fi {vkl |µi (k), σi (k)} (6.2.10)
l=1 i=1

obeying the constraint in (6.2.3). If λ be a Lagrange multiplier, we get,

n
X m
X m
X
Ω(k) = loge pi (k)fi {vkl |µi (k), σi (k)} − λ{ pi (k) − 1} (6.2.11)
l=1 i=1 i=1
82

The mean µi (k) of the Gaussian components for which Ω(k) in (6.2.11) would be

maximum can be calculated from

∂Ω(k)
=0 (6.2.12)
∂µi (k)

Now
∂ −(v l −µ (k))2
Xn pi (k) [ √ 1 2 exp{ 2σ k
2
i
}]
∂Ω(k) ∂µi (k) 2πσi (k) i (k)
= Pm l
∂µi (k) l=1 i=1 pi (k)fi {vk |µi (k), σi (k)}
n
X p (k)fi {vkl |µi (k), σi (k)} {vkl − µi (k)}
= Pmi l
.
l=1 i=1 pi (k)fi {vk |µi (k), σi (k)} σi (k)

Using (6.2.8) and (6.2.12), we get

n
X {vkl − µi (k)}
P {fi (vkl )|vkl } = 0.
l=1
σi (k)

Therefore,
Pn
P {fi (vkl )|vkl }vkl
µi (k) = Pl=1
n l l
. (6.2.13)
l=1 P {fi (vk )|vk }

Similarly, new estimates for variance σi2 (k) can be obtained by putting

∂Ω(k)
=0 (6.2.14)
∂σi (k)

Solving (6.2.14), gives

Pn
P {fi (vkl )|vkl }{vkl − µi (k)}2
σi2 (k) = l=1
Pn l l
(6.2.15)
l=1 P {fi (vk )|vk }

The estimates for the weights pi (k) of the Gaussian components densities, are found

from:
∂Ω(k)
=0 (6.2.16)
∂pi (k)
n
X fi {vkl |µi (k), σi (k)}
Pm l
−λ=0 (6.2.17)
l=1 i=1 pi (k)fi {vk |µi (k), σi (k)}
 83

Using (6.2.8) and (6.2.17) we get

n
X P {fi (v l )|v l }
k k
−λ=0
l=1
pi (k)
n
1X
⇒ pi (k) = P {fi (vkl )|vkl } (6.2.18)
λ l=1
Also, from (6.2.3) and (6.2.18), we can write
m
X Xm n
1X
pi = P {fi (vkl )|vkl }
i=1 i=1
λ l=1
= 1

Therefore
n X
X m
λ= P {fi (vkl )|vkl } (6.2.19)
l=1 i=1

Now, using (6.2.9) and (6.2.19), we get

λ=n

Inserting λ into (6.2.18), gives

n
1X
pi (k) = P {fi (vkl )|vkl }. (6.2.20)
n l=1

Thus, (6.2.13), (6.2.15) and (6.2.20) respectively give the a posteriori updated values

of the means, variances and weights of the Gaussian components. Inserting the cor-

responding values of the means, variances and weights in (6.2.2), the resulting PDF

(observed in the simulation results in Section 6.4) closely approximates the actual

non-Gaussian noise PDF.

In the proposed algorithm, the parameters (means, variances and weights) have been

evaluated in two stages - stage 1: a priori update and stage 2: a posteriori update.

At the start of each time step, the parameters of the i-th Gaussian component are
84

set by a newly proposed a priori update. Next, in the a posteriori update, the final

values of the parameters are evaluated by (6.2.13), (6.2.15) and (6.2.20). In the a

priori update, the mean of the individual Gaussian models at k-th time step are set

such that they are symmetrically distributed about the a posteriori resultant mean

of the previous time step i.e. µ(k − 1). By this strategy an asymmetrical and multi

modal distributed non-Gaussian noise can be characterized too.

2
The a priori variance σ0i (k), a priori mean µ0i (k) and a priori weight p0i (k) are

selected as:

The a priori weight:

1
p0i (k) = (6.2.21)
m
where m is the number of Gaussian components.

The a priori variance:

2 2
σ0i (k) = σ0i (k − 1) (6.2.22)

The a priori mean µ0i (k) is set as

m+1 1
µ0i (k) = µ(k − 1)(1 + ( − i) ) (6.2.23)
2 m

when m is an odd number.

m
When m is an even number, for i = 1 to 2
,

m+2 1
µ0i (k) = µ(k − 1)(1 + ( − i) ), (6.2.24)
2 m
m
and, for i = 2
+ 1 to m,

m 1
µ0i (k) = µ(k − 1)(1 + ( − i) ), (6.2.25)
2 m

where,

µ(k − 1) = p1 (k − 1)µ1 (k − 1) + ... + pm (k − 1)µm (k − 1), (6.2.26)

 85

is the resultant a posteriori mean of n number of data set at the (k − 1)-th time

instant.

6.3 Algorithm

In this section the algorithm for parameter estimation of a NSNGN is summarized

below

1. At the start of the k-th time step, the a priori parameters, i.e., weight p0i (k),
2
mean µ0i (k) and variance σ0i (k) of the i-th Gaussian component are set as:

• a priori weight of Gaussian components (by (6.2.21)),

1
p0i (k) = m

2
• a priori variance (by (6.2.22)), σ0i (k) = σi2 (k − 1)

• a priori mean (by (6.2.23)-(6.2.26)),

When m is an odd number,

µ0i (k) = µ(k − 1)(1 + ( m+1

2
− i) m1 ).

When m is an even number,

µ0i (k) = µ(k − 1)(1 + ( m+2

2
− i) m1 ),

for i = 1 to m/2, and,

µ0i (k) = µ(k − 1)(1 + ( m2 − i) m1 ),

for i = m/2 + 1 to m.

2. The posterior probability of the i-th Gaussian component (6.2.8) using the a

priori values in step 1 is given by,

p0i (k)fi {vkl |µ0i (k), σ0i (k)}

P {fi (vkl )|vkl } = Pm l
j=1 p0j (k)fj {vk |µ0j (k), σ0j (k)}
86

3. The a posteriori values of the parameters of the Gaussian components are esti-

mated by (6.2.13), (6.2.15) and (6.2.20) as:

a posteriori weight of the Gaussian component,

n
1X
pi (k) = P {fi (vkl )|vkl }
n i=1
a posteriori mean
Pn
P {fi (vkl )|vkl }vkl
µi (k) = Pl=1
n l l
l=1 P {fi (vk )|vk }

a posteriori variance
Pn
P {fi (vkl )|vkl }{vkl − µi (k)}2
σi2 (k) = l=1
Pn l l
l=1 P {fi (vk )|vk }

4. a posteriori resultant mean of the NSNGN

µ(k) = p1 (k)µ1 (k) + ... + pm (k)µm (k)

6.4 Simulation Results

To validate the algorithm for PDF estimation of a NSNGN, the following three exam-

ples are simulated. In Example I a non-stationary asymmetrical Rayleigh distributed

noise (non zero mean) has been taken as the test signal. Example II is taken from a

real world situation where a non-stationary glint noise is observed in a radar tracking

system [95]-[96]. The distribution of glint noise is non-Gaussian and heavy tailed.

To test the validity of the proposed algorithm in case of multi-modal distribution,

a bi-modal distributed stationary data has been considered in Example III. And in

this case (Example III ) the performance of the proposed algorithm has been com-

pared with an existing expectation maximization (EM) method (which is usually used

for parameters estimation of a Gaussian mixture model [86] where the parameters -

means, variances and weighting co-efficients are assumed to be fixed.)

 87

6.4.1 Example I

In this example a random sequence of test noise signal is generated from a Rayleigh

distribution with linearly varying parameter b from 1 to 3 during 50 time samples.

The algorithm is used to estimate the parameters of this noise sequence at each time

instant. The equivalent mean and variance of the noise distribution varies from 1.25 to

3.76 and 0.43 to 3.86 respectively. A sequence of 1000 individual data sets are taken.

The number of Gaussian densities in the mixture is fixed to m = 5. The values of the

parameters at the initial time step are taken as: the resultant a posteriori mean µ(1) =

1.2, a priori standard deviations σ01 (1) = 0.45, σ02 (1) = 0.4, σ03 (1) = 0.3, σ044 (1) =

0.35, σ05 (1) = 0.25 and a priori weights p01 (1) =p02 (1) =p03 (1) =p04 (1) =p05 (1) =

0.2.

Figure 6.1: Estimated PDF at 5th time step

The estimated noise PDFs at the 5th, 25th and 45th time steps are shown in

Figs. 6.1-6.3 respectively. It may be seen that in each case the resulting PDF (red

one) closely resembles the actual PDF (blue one) of the noise while the black one is the

Gaussian PDF with mean and covariance of the corresponding Rayleigh distribution.
88

Figure 6.2: Estimated PDF at 25th time step

Figure 6.3: Estimated PDF at 45th time step

 89

6.4.2 Example II

In some radar tracking systems [96] there exists a non-stationary heavy tailed non-

Gaussian noise known as glint noise. The glint noise (Fig. 6.4) PDF is usually modeled

by a student t distribution [95]. Here, the simulation uses a student t distributed

0
Glint noise

−5

−10

−15

−20
0 10 20 30 40 50
Time samples

Figure 6.4: Plot of non-stationary glint noise

noise for 50 time samples. The noise is made non-stationary by linearly varying the

degree of freedom from 6 to 2.1 during 50 time samples. The equivalent mean of

the noise distribution is zero and the variance varies from 1.5 to 21. The proposed

algorithm is illustrated by estimating the parameters of this NSNGN sequence at

each time instant. A sequence of 1000 individual data sets are taken and processed

with the proposed algorithm. The number of Gaussian densities in the mixture is

fixed to m = 5. The initial a priori values of the parameters are selected as: weights
2 2
p01 (1) =p02 (1) =p03 (1) =p04 (1) =p05 (1) = 0.2, variances σ01 (1) = 0.02, σ02 (1) = 2,
2 2 2
σ03 (1) = 0.15, σ04 (1) = 1.5, σ05 (1) = 10 and the initial value of a posteriori resultant
90

Figure 6.5: Estimated PDF at 5th time step

mean µ(1) = 0. The estimated noise PDFs at the 5th, 25th and 45th time steps

are shown in Figs. 6.5-6.7 respectively. It may be observed that in each case the

resulting PDF (red one) closely resembles the actual PDF (blue one) while the black

one is the Gaussian PDF with mean and covariance of the corresponding student t

distribution.

6.4.3 Example III

In many situations the noise PDF is bi-modal or multi-modal instead of being uni-

modal. The proposed algorithm is tested here and compared with an EM algorithm

[86] to estimate the parameters of the PDF of a bi-modal distributed data set. Here

a sequence of 1000 data set are taken. The number of Gaussian densities in the

mixture are fixed to m = 5. Initially the values are selected as: resultant mean

µ = 1.5, standard deviations σ01 = σ02 = σ03 = σ04 = σ05 = 0.3 and weights

p01 =p02 =p03 =p04 =p05 = 0.2. Same initial values are selected for both the EM

and the proposed algorithms. Starting from the initial values of the parameters, the
 91

Figure 6.6: Estimated PDF at 25th time step

Figure 6.7: Estimated PDF at 45th time step

92

EM algorithm iterates until the sequence of estimates of the parameters converge to

a stationary point of the likelihood function [97]. The histogram of the data and the

estimated PDF are shown in Figs. 6.8 and 6.9 respectively.

Histogram
12

10

0
−3 −2 −1 0 1 2 3 4 5

Figure 6.8: Histogram of bi-modal distributed data

0.4
proposed algorithm
0.35 EM algorithm
probability density

0.3

0.25

0.2

0.15

0.1

0.05

0
−4 −3 −2 −1 0 1 2 3 4 5
pdf argument

Figure 6.9: Estimated PDF

It is observed that resultant PDF (Fig. 6.9) using the proposed algorithm closely

characterizes the bi-modal distributed data (Fig. 6.8), whereas the EM algorithm fails

to do so.
 93

6.5 Conclusion

In this chapter an algorithm is proposed to estimate the parameters of the PDF of a

noise signal, which may be stationary or non-stationary, Gaussian or non-Gaussian,

zero mean or non-zero mean and uni-modal or multi-modal. Three examples illus-

trate the effectiveness of the new algorithm. In the first example, the PDF of a

non-stationary (non-zero mean) Rayleigh distributed noise has been estimated. In

the second, a real world example is taken where the noise is a non-stationary (zero

mean) glint noise. The proposed algorithm is also tested for a bi-modal distributed

stationary data in the third example and compared with the EM algorithm. In all the

cases, the estimated PDFs closely follow the actual ones of the test noise sequence.

So the algorithm seems to be capable of estimating the PDF of an NSNGN. During

estimation of noisy signals, if the noise model (obtained through the proposed tech-

nique) is suitably included in filters, it is expected to increase the filtering efficiency.

The proposed algorithm is applicable in practical situations where it is possible to

acquire noise data beforehand (at least through simulation). The importance of the

algorithm lies in the fact that in many cases the noise is found to be non-stationary

and/or non-Gaussian and/or multi-modal. In those cases the present work has been

proven to be useful.
94

.
Chapter 7

Likelihood Function Modeling of

Particle Filter in Presence of
Non-Stationary Non-Gaussian
Measurement Noise

A generalized likelihood function model of an sampling importance resampling (SIR)

particle filter(PF) has been proposed for state estimation of a nonlinear system in

presence of non-stationary, non-Gaussian white measurement noise. The measure-

ment noise is modeled by Gaussian mixture probability density functions and the

noise parameters are estimated by maximizing the log likelihood function of the noise

model. This model is then included in the likelihood function of the SIR particle filter

(PF) at each time step for online state estimation of the system. The performance of

the proposed algorithm has been evaluated by estimating the states of i) non-linear

system in presence of non-stationary Rayleigh distributed noise and ii) a radar track-

ing system in presence of glint noise. The simulation results show that the proposed

modified SIR PF offers best performance among the other algorithms considered for

these examples.

95
96

7.1 Introduction

Optimal solution to the state estimation problem of nonlinear systems is still a re-

search problem - especially in the presence of measurement noise which is neither

Gaussian in distribution and nor stationary. The Particle filter, (also known as se-

quential importance (re-)sampling (SIS/SIR) and sequential Monte Carlo methods

(SMS)) [17]-[21] has proven its worth in case of dynamic state estimation of nonlinear

systems fed by non-Gaussian noise. It too yields a suboptimal solution. Moreover it

works on the assumption that the measurement noise is a zero mean white sequence

with a known probability density function (PDF) [21]. The present work proposes an

improvement in some of these issues.

The observation likelihood function, responsible for determining the particles

weights, depends on measurement noise. The particle weights in turn determine

how they are resampled. The likelihood function in this way has strong influence on

the quality of estimation. The paper [98], contains experimental results which show

a large influence of observation likelihood function on tracking performance of three

different tracking tasks for different parameter values of an assumed measurement

model. However not much work is reported in the literature on modeling of the ob-

servation likelihood function which is necessary to calculate the particle weights from

observed measurements. Often no details are given as to how the observation like-

lihood function model of a particle filtering algorithm is determined. In the present

work a new scheme is proposed for dynamic state estimation of a nonlinear system

using a particle filter in presence of non-stationary non-Gaussian measurement noise.

Non-Gaussian densities of noise are closely approximated by Gaussian mixture

models [82] - [84]. In [99], the author discussed that a posteriori PDF from the
 97

Bayesian recursion relation can be approximated by a weighted sum of Gaussian

probability density functions. In [100], the author proposed a special sequential

Monte-Carlo method, the mixture Kalman filter, which uses random mixture of nor-

mal distributions to represent target distributions. This technique is some time also

referred to as Rao-Blacwellization [101] and marginalized particle filter [102]. These

filters yet were developed for linear or nonlinear models on the assumption of zero

mean white Gaussian process and measurement noise. The paper [103], investigated

Bayes optimal adaptive estimation for a linear discrete time system with unknown

Markovian state space and measurement noise. For a nonlinear system, the paper

[104], developed algorithm of several Gaussian sum particle filters for a state space

model with non-Gaussian noise. These filters, approximating the filtering predictive

distribution by weighted Gaussian mixture are basically banks of Gaussian particle

filters [105]. None of these papers, however, were developed for state estimation of

nonlinear system on the assumption of non-Gaussian non-stationary measurement

noise. Moreover in case of Gaussian mixture filters e.g. Gaussian sum particle filter

[104]- bank of a Gaussian particle filters (GPF) run parallely, similarly in case of

mixture Kalman filter [100] - bank of a Kalman filters runs parallely so the compu-

tational load will be higher than the new SIR PF where a single PF runs to estimate

the states in presence of non-Gaussian noise.

As the likelihood function depends on the measurement noise distribution, first pa-

rameter estimation and modeling of the noise has been carried out. The modeling and

parameters (mean, variance and weighting coefficient) estimation of non-stationary

non-Gaussian noise (NSNGN) has been describe in detail in the Chapter 6, (Sec. 6.2-

Sec. 6.3) [106]. The present Chapter seeks an improvement to the particle filter
98

algorithm [21] by proposing a generalized observation likelihood model through mea-

surement noise preprocessing. The noise is modeled using Gaussian mixture PDFs

and additionally, the noise parameters are evaluated by maximizing the log likeli-

hood function of a data set of some independent noise sequences. This model and

evaluated parameters are used in the likelihood function of the SIR particle filter for

state estimation of a nonlinear system in presence of zero mean or non-zero mean

NSNGN sequence. It is a generalized noise model which can be used for PDF es-

timation of stationary or non-stationary, Gaussian or non-Gaussian, zero mean or

non-zero mean, unimodal or multi-modal noise sequence. This modified SIR particle

filter can be used for state estimation of a nonlinear system in presence of any type of

white measurement noise. This chapter is organized as follows. Sec. 7.2, recalls the

basic SIR particle filter algorithm. Sec. 7.3 shows how to include the proposed model

of the likelihood function in the SIR filter. The model order growth over iteration

of the Gaussian mixture models for non-Gaussian measurement noise is discussed in

Sec. 7.4. Simulation results are shown in Sec. 7.5. Finally conclusions are drawn in

Sec. 7.6.

7.2 SIR Particle Filter

The SIR filter proposed in [21] is a Monte Carlo method that can be applied to

recursive Bayesian filtering problems. The state vector dynamics xk ∈ <n is assumed

to evolve according to following system model.

xk = fk−1 (xk−1 ) + wk−1 (7.2.1)

 99

wk is a zero mean, white noise sequence. The PDF of wk is assumed to be known.

The measurements yk related to the state vector via the measurement equation

yk = hk (xk ) + vk (7.2.2)

vk is another zero mean, white noise sequence of known PDF and independent of

process noise. SIR algorithm needs samples (particles) Σk−1 = {xk−1 (j) : j = 1, ..., n}

from p(xk |xk−1 (j)). The SIR filter propagates the random sample set by the following

steps.

• Prediction: Each sample in Σk−1 is passed through the state equation (7.2.1)

to obtain samples from the prior density at time step k

x∗k (j) = fk−1 (xk−1 (j)) + wk−1 (j) (7.2.3)

where wk is a sample drawn from the PDF of process noise p(wk ) .

• Update: i) Calculate the likelihood function p(zk |x∗k (j)) for each sample in the

set Σ∗k = x∗k (j) : j = 1, ..., n . ii) Calculate the normalized weights for each

samples
p(zk |x∗k (j))
qk (j) = n (7.2.4)
X
p(zk |x∗k (j))
j=1

The weights qk (j) represent the probability mass associated with element j of

Σ∗k .

• Resample n times from the discrete distribution defined by Σ∗k and qk = {xk (i) :

j = 1, ..., n} to generate samples Σk = {x∗k (j) : j = 1, ..., n} so that for any j 0 ,

P r{xk (j 0 ) = x∗k (j)} = qk (j)

100

7.3 Likelihood Function Modeling of an SIR Par-

ticle Filter

The likelihood function p(zk |xk ) needs to be available in (7.2.4) for point wise evalua-

tion. In practice the likelihood function is evaluated as p(zk |xk ) = pN (zk ; hk (xk ), Rk )

assuming vk to be always a zero mean white Gaussian noise with covariance Rk . When

vk turns out to be non stationary and non-Gaussian with a non-zero mean, this is

expected to lead to introduce gross errors in the estimation process.

In case the noise sequence is non stationary and non-Gaussian with a non-zero

mean, the PDF may be approximated using the newly proposed algorithm in Sec. 6.3

and subsequently the likelihood function can be evaluated as follows:

Let x and y be two independent n vector valued random variables, and z is defined

as their sum.

z = x + y; (7.3.1)

It directly follows that the conditional PDF may be found as [16]

pz|x (z|x) = py (z − x) (7.3.2)

Thereby from (7.2.2) we can write

p(zk |xk ) = p(zk − hk (xk ))

= p(vk ) (7.3.3)

Following these lines, in the SIR algorithm, the likelihood function (in (7.2.4)) can

be evaluated using (6.2.2). Therefore the modified SIR algorithm takes the following

form including measurement noise preprocessing (using Algorithm in Sec. 6.3),

 101

7.3.1 Algorithm: New SIR particle Filter

[{xk (j), qk (j)}nj=1 ] = SIR[{xk−1 (j), qk−1 (j)}nj=1 , zk ]

• for j = 1 : n

– Draw

xk (j) ∼ p(xk |xk−1 (j)) (7.3.4)

– calculate

qk (j) = p(zk |xk (j)) (7.3.5)

= p(vk (j))
Xm
pi (k) {vk (j) − µi (k)}2
= 1/2 σ (k)
exp[− 2
]
i=0
(2π) i σ i (k)

(using 6.2.1, 6.2.2 & 7.3.3)

• end for

• Calculate total weight: t = SUM[{qk (j)}nj=1 ]

• for j = 1 : n

– Normalized: qk (j) = t−1 qk (j)

• end for

• Resampling (procedure explained in [17, 18])

– [{xk (j), qk (j)}nj=1 ] = RESAMPLE[{xk (j), qk (j)}nj=1 ]

102

7.4 Model Order Growth over Iteration for Gaus-

sian Mixture Models of Non-Gaussian Mea-
surement Noise

In this section the order of the mixture models over iteration on general Bayesian

frame work is discussed. Consider the system given in Section 7.2 ,(7.2.1 - 7.2.2), the

non-Gaussian noise in the measurement (7.2.2) is approximated by a finite number

of Gaussian mixture model (6.2.1).

The measurement noise PDF is therefore given by

m
X
p(vk ) = pi (k)N (vk ; µi (k); σi2 (k)) (7.4.1)
i=1

We assume that the process noise is Gaussian. The prediction density of the state at

time k via the Chapman-Kolmogorov equation is

Z
p(xk |y1:k−1 ) = p(xk |xk−1 )p(xk−1 |y1:k−1 )dxk−1 (7.4.2)

At time step k, when a measurement yk becomes available, the update of the predic-

tion PDF via the Bayes rule is given by

p(xk |y1:k ) = Ck p(yk |xk )p(xk |y1:k−1 ) (7.4.3)

Where Ck is the normalized constant

Z
Ck = p(yk |xk )p(xk |y1:k−1 )dxk (7.4.4)

Consider k = 1, after the arrival of y1 , the posterior density can be expressed as

p(x1 |y1 ) = C1 p(y1 |x1 )p(x1 ) (7.4.5)

 103

Now,

p(y1 |x1 ) ≈ p(v1 )

Xm
= pi (1)N (v1 ; µi (1); σi2 (1))
i=1
m
X
p(y1 |x1 ) = pi (1)pi (y1 |x1 ) (7.4.6)
i=1

where, pi (y1 |x1 ) = N (v1 ; µi (1); σi2 (1)). Therefore substituting (7.4.6) into (7.4.5) we

get,
m
X
p(x1 |y1 ) = C1 pi (1)pi (y1 |x1 )p(x1 )
i=1
m
X
or, p(x1 |y1 ) = αi (1)e
pi (x1 |y1 ) (7.4.7)
i=1

where pei (x1 |y1 ) ∝ pi (y1 |x1 )p(x1 ) and αi (1) is the normalized weight. Again, after

arrival of y2 , the posterior density is

p(x2 |y1:2 ) = C2 p(y2 |x2 )p(x2 |y1 ) (7.4.8)

Xm
= C2 pi (2)pi (y2 |x2 )p(x2 |y1 ) (7.4.9)
i=1

Now, using (7.4.2) we can write,

Z
p(x2 |y1 ) = p(x2 |x1 )p(x1 |y1 )dx1 (7.4.10)

Substituting (7.4.7) into (7.4.10) we get,

Z m
X
p(x2 |y1 ) = p(x2 |x1 ) αi (1)e
pi (x1 |y1 )dx1
i=1
m
X Z
= αi (1) pei (x1 |y1 )p(x2 |x1 )dx1
i=1
m
X
or, p(x2 |y1 ) = αi (1)e
pi (x2 |y1 ) (7.4.11)
i=1
104

Now, replacing (7.4.11) into (7.4.9) we get,

m
X m
X
p(x2 |y1:2 ) = C2 pi (2)pi (y2 |x2 ) αj (1)e
pj (x2 |y1 )
i=1 j=1
Xm Xm
= C2 αj (1)pi (2)pi (y2 |x2 )e
pj (x2 |y1 )
i=1 j=1
m 2
X
= αj (2)pj (y2 |x2 )e
pj (x2 |y1 )
j=1
m 2
X
p(x2 |y1:2 ) = αj (2)e
pj (x2 |y1:2 ) (7.4.12)
j=1

where pej (x2 |y1:2 ) = pj (y2 |x2 )e

pj (x2 |y1 ). Similarly after arrival of k-th measurement,

the posterior PDF is

m k
X
p(xk |y1:k ) = αj (k)e
pj (xk |y1:k ) (7.4.13)
j=1
So we can see that the mixture model order increases exponentially on the arrival

of a new measurement. In practical applications updating with an exponentially

growing mixture model order is extremely difficult to handle. But this situation does

not arise in case of an SIR particle filter. From the above derivation we can see

that the p(xk |y1:k−1 ) term is the responsible for the model order growth. But in SIR

particle filter, predicted PDF p(xk |y1:k−1 ) is approximated by the term p(xk |x1:k−1 )

[17]-[21] (i.e. p(xk |y1:k−1 ) ≈ p(xk |x1:k−1 )). which is independent of the measurement.

Therefore the posterior PDF in (7.4.3) at the k-th time step can be expressed using

(6.2.1) as

p(xk |y1:k ) = Ck p(yk |xk )p(xk |y1:k−1 ) (7.4.14)

Xm
≈ Ck pi (k)pi (yk |xk )p(xk |x1:k−1 ) (7.4.15)
i=1
From the SIR PF algorithm in Sec. 7.3.1, we can also see that in the measurement

update stage there is no need to pass the importance weights from one time step of
 105

the algorithm to the next. From (7.3.5), we can say that the importance weights

are independent of the previous measurements. And in time update stage (7.3.4),

importance sampling density for the SIR PF is independent of measurements. So

the mixture model order for non-Gaussian measurement noise would not grow over

iteration in case of an SIR PF.

7.5 Simulation Results

The new SIR PF proposed in this chapter is applied here to two different systems.

In Example I, a simple nonlinear time series model is chosen which has been used

extensively in the literature for benchmarking numerical filtering techniques [21], [18].

Here the measurement noise is considered to be a non-stationary Rayleigh distributed

noise. And in Example II, the bearing only tracking model [21], which is of interest in

defence applications, is presented in presence of glint noise. The glint noise is a heavy

tailed non-Gaussian distributed noise [107]-[65]. The performance of the new SIR PF

was compared with a conventional SIR PF and a mixture Kalman filter (MKF) [100].

7.5.1 Example I

Here the modified SIR filter is used to estimate the states of a nonlinear system in pres-

ence of a non-stationary non-Gaussian non-zero mean measurement noise sequence

and the results are compared with that of a usual SIR PF, MKF [100] and extended

Kalman filter (EKF) [14]-[15]. Consider the nonlinear model [21] given below:

25xk−1
xk = 0.5xk−1 + + 8cos(1.2(k − 1)) + wk (7.5.1)
[1 + xk−1 ]2
x2k
yk = + vk (7.5.2)
20
106

RMSE with 500 particles of 100 Monte carlo run

90
SIR PF
80 New SIR PF
MKF
KF
70

60
RMSE

50

40

30

20

10

0
0 10 20 30 40 50
Time

Figure 7.1: RMSE of 100 Monte Carlo runs

where, wk is a zero mean Gaussian white noise with variance 10, whereas, vk is

assumed to be a non-stationary and non-Gaussian - a non stationary Rayleigh dis-

tributed signal. The simulation was carried out for 50 time samples. The sequence

of noise was generated randomly from a Rayleigh distribution with linearly varying

parameter b from 1 to 2 during 50 time samples. The equivalent variance of the noise

distribution varies from 0.43 to 1.72. The Gaussian mix maximum likelihood algo-

rithm (Sec. 6.2, 6.3) has been used to estimate the parameters of the non-stationary

non-Gaussian noise sequence at each time step. A sequence of 1000 individual data

set were taken and processed as shown in Chapter 6, Sec. 6.4.1. The number of

Gaussian densities in the mixture was fixed to m = 5. The initial value of weighting
1
co-efficients pi (1) is set at m
. In this simulation initial value of the resultant mean
2 2
was taken as µ(1) = 1.2 and variance was σ01 = .. = σ05 = .252 .

The resulting PDF using the algorithm in Sec. 6.3, closely estimates the actual
 107

Comparison of MSE w.r.t number of particle

700
New SIR
SIR
600
MKF
KF

Mean Square Error

500

400

300

200

100

0
1 2 3
Number of particles: 500 − 100 − 50

Figure 7.2: Average of MSE over time samples with number of particles 50, 100 and
500 (for 100 Monte Carlo runs).

PDF of the non-stationary Rayleigh distributed noise present in the measurement

(similarly as shown in Figs. 6.1 - 6.3). This resulting noise PDF is included in the

likelihood function of the SIR algorithm in (7.2.4). The performance of the modified

SIR PF has been compared with the conventional SIR PF, MKF and EKF. In the

conventional SIR filter the likelihood function has been taken as a Gaussian one,

p(zk |xk ) = 1
(2πRk )1/2
exp[− yh −h k (xk )
2Rk
], where Rk is the time variant noise covariance

which varies from 0.43 to 1.72 during the 50 time samples. The same measurement

noise covariance R(k) is selected for EKF. In case of the MKF [100],[109], here vk ∼

N (0, Rk (Λk )). The indicator vector Λk is a discrete latent variable which takes an

integer value between 1 ∼ m(number of models). The measurement noise vk is

P
approximated by m 2
i=1 pi N (0, σi ). The indicator random variable:
108

1, if vk ∼ N (0, σ12 )
n 2, if v ∼ N (0, σ 2 )
k 2
Λk , . with P (Λk = 1) = p1 , ....,P (Λk = m) = pm . The
....
2
m, if vk ∼ N (0, σm )
number of models M = 5 was selected in the Mixture Kalman filter with five different

values of measurement noise covariance σ(i)2 = 1.2i−1 × σ 2 , where i = 1, .., M and

σ 2 = .43 was selected. Since the state space and measurement model is nonlinear, it

must be linearized before applying the MKF. For all the filters, the initial value of

mean and state covariance was selected as 0.1 and 5 respectively. The value of the

process noise covariance is Q = 10 for all the filters. The root mean square error

(RMSE) of each filter was computed. Fig 7.1 shows RMSE of each filter with 500

particles for L = 100 Monte Carlo runs. The RMSE of xk is defined as:
v
u L
u1 X
RM SExk = t × (xk,j − x̂k,j ) (7.5.3)
L j=1

The comparison of average MSE (where MSE = RMSE2 ) over the 50 time samples

among the filters in terms of the number of particles is shown in Fig. 7.2. The

average MSE was compared for the number of particles 50, 100 and 500. Even for

small number of particle new SIR PF performs significantly better than EKF. An

increase in the number of particles reduces the MSE further for new SIR PF, SIR PF

and MKF. The performance of the new SIR PF becomes very close for the number

of particles 500 and 100. The performance of the new SIR PF appears to be better

than the other filters taken for comparison.

7.5.2 Example II

Here radar target tracking in presence of glint noise is considered. The performance

of the new SIR PF is compared to the SIR PF and MKF. Radar tracking is an
 109

important area of research in defence applications. Practical radar tracking systems

are rarely Gaussian due to many factors. One of them is changes in the aspect towards

the radar, which can cause irregular electromagnetic wave reflection, resulting in

significant variations of radar reflections [65]. This phenomenon gives rise to outliers

in angle tracking, and it is referred to as target glint. It is found that glint has a

long-tailed PDF [65]. The glint has been modeled by a Students t distribution with

two degree of freedom [64]. For student t distribution (with 2 degree of freedom),

the estimated PDF using the proposed maximum log likelihood algorithm (Sec. 6.2,

Algorithm) for noise modeling is shown in Fig. 7.3 (where m = 5, the initial value of

weighting co-efficients pi (1) = 1/m, initial value of resultant mean µ(1) = 0, variances
2 2 2 2 2
σ01 (1) = 0.02, σ02 (1) = 2, σ03 (1) = 0.15, σ04 (1) = 1.5, σ05 (1) = 10). A mixture model

of two zero-mean Gaussian PDFs has been proposed for glint noise in [65], based

on the statistical analysis. This model consists of one Gaussian PDF with a high

probability and small variance and another with small probability of occurrence and

very high variance. In [110], the glint noise was alternatively modeled by a mixture

of zero-mean, small-variance Gaussian PDF with heavy-tailed Laplacian distribution.

The last two models are most commonly used for glint noise modeling [111].

Considering the target moves on an x-y plane (2-D space) according to following

process model [21]:

   
2
1 T 0 0 T /2 0
   
 0 1 0 0   T 0 
   
Xk =   Xk−1 +   wk−1 (7.5.4)
 0 0 1 T   0 T 2 /2 
   
0 0 0 1 0 0

and measurement model:

Zk = tan−1 (yk /xk ) + vk (7.5.5)

110

Estimated PDF
0.4
Student T pdf
Estimated pdf
0.35

0.3
probability density

0.25

0.2

0.15

0.1

0.05

0
−15 −10 −5 0 5 10 15 20 25
pdf argument

Figure 7.3: Estimated PDF of student t distribution

h i
where Xk = xk vxk yk vyk , xk and yk denote the cartesian co-ordinate position

of the target, vxk and vyk denote the velocities in x and y directions respectively. wk−1

is a zero mean white Gaussian noise with covariance Q, where Q = qI2 , I2 is a 2 × 2

√
identity matrix and q = .001. The measurement noise vk was considered to be a

glint noise which is modeled by following assumptions [107], [108].

p(vk ) = (1 − ²)N (vk ; 0; σr2 ) + ²N (vk ; 0; κσr2 ) (7.5.6)

where glint probability ² = .1, κ = 1000 and the measurement noise covariance R =

σr2 = .0052 The target was simulated for 50 time samples with sampling interval T = 1
h iT
sec and the initial value of state is selected as x0 = −0.05, 0.001, 0.7, −0.055 .

The performance of the new SIR PF has been compared with the conventional

SIR PF and MKF. The initial prior distribution of the state vector for all the
h iT
filters was set with a mean x0 = 0, 0, 0.4, −0.05 , and covariance P0 =
 111

 
0.52 0 0 0
 
 0 0.0052 0 0 
 
 . In the conventional SIR filter, the measurement noise
 0 0 0.32 0 
 
0 0 0 0.012
covariance is R = 0.0052 . In case of the MKF, here vk ∼ N (0, Rk (Λk )). The in-

dicator vector Λk is a discrete latent variable which takes an integer value between

1 ∼ M (number of models). The number of models, M = 5, is selected in the

Mixture Kalman filter with five different values of measurement noise covariance

R(i) = 1000i−1 × R, where i = 1, .., M and R = 0.0052 is selected. Since the mea-

surement model is nonlinear, it must be linearized before applying MKF. The root

mean square error (RMSE) of the state vector for each filter was computed. RMSE

of position and velocity along x-axis and y-axis using 1000 particles for L = 100

Monte Carlo runs are shown in Figs 7.4-7.7. The comparison of average MSE over

the 50 time samples among the filters in terms of the number of particles is shown in

Figs. 7.8-7.9.

The average MSE was compared for the number of particles 5000, 1000, 500 and

100. Even for a small number of particles the new SIR PF performs significantly

better than the other filters here. An increase in the number of particles reduces

the MSE further for the new SIR PF, SIR PF and MKF. The performance of the

new SIR PF becomes very close for the number of particles 5000, 1000 and 500. The

simulation results show that the new SIR PF gives considerably better results also

with a small number of particles since the importance weight of SIR PF depends on

the likelihood function i.e. measurement noise PDF which is estimated very closely by

the algorithm in Sec. 6.2. Among all the filters taken for comparison the performance

of the new SIR PF is much better.

112

RMSE with 1000 samples of 100 Monte carlo runs

0.09
SIR
0.08 New SIR
RMSE of velocity along x−axis

MKF
0.07

0.06

0.05

0.04

0.03

0.02

0.01

0
0 10 20 30 40 50
Time

Figure 7.4: RMSE of 100 Monte Carlo runs

RMSE with 1000 samples of 100 Monte carlo runs

SIR
0.12
New SIR
RMSE of velocity along y−axis

MKF

0.1

0.08

0.06

0.04

0.02

0
0 10 20 30 40 50
Time

Figure 7.5: RMSE of 100 Monte Carlo runs

 113

RMSE with 1000 samples of 100 Monte carlo runs

4.5
SIR
4 New SIR

RMSE of position along x−axis

MKF
3.5

2.5

1.5

0.5

0
0 10 20 30 40 50
Time

Figure 7.6: RMSE of 100 Monte Carlo runs

RMSE with 1000 samples of 100 Monte carlo runs

6
SIR
New SIR
RMSE of position along y−axis

MKF
5

0
0 10 20 30 40 50
Time

Figure 7.7: RMSE of 100 Monte Carlo runs

114

Comparison of MSE w.r.t number of particle

0.12
SIR
0.1 MKF
RMSE of vy

0.08 new SIR

0.06
0.04
0.02

1 2 3 4
Number of particles: 5000 − 1000 − 500 − 100

0.08
SIR
0.06 MKF
RMSE of vx

new SIR
0.04

0.02

0
1 2 3 4
Number of particles: 5000 − 1000 − 500 − 100

Figure 7.8: Comparison of average MSE of velocity along x-axis and y-axis for the
number of particles: 5000 - 1000 - 500 - 100

Comparison of MSE w.r.t number of particle

3
SIR
MKF
RMSE of y

2 new SIR

0
1 2 3 4
Number of particles: 5000 − 1000 − 500 − 100

2.5
SIR
2 MKF
RMSE of x

new SIR
1.5

0.5

0
1 2 3 4
Number of particles: 5000 − 1000 − 500 − 100

Figure 7.9: Comparison of average MSE of position along x-axis and y-axis for the
number of particles: 5000 - 1000 - 500 - 100
 115

7.6 Conclusion

In this chapter, a generalized likelihood function has been derived for an SIR particle

filter. The likelihood function is evaluated from measurement noise history using

Gaussian mixture models. The parameters of the noise are estimated by maximizing

log likelihood function. This model and the estimated parameters are included in

the likelihood function of the SIR particle filter. This is a generalized model as it

can represent the PDF of noise sequence be it Gaussian/non-Gaussian, zero mean or

non-zero mean, stationary/non-stationary. In this way the new SIR particle filter is

able to estimate states in presence of any type of white measurement noise sequence.

And since the importance weights of a SIR PF completely depend on the likelihood

function, it performs considerably better with a small number of particles than the

conventional SIR PF and MKF as long as the trend of measurement noise parameters

do not change drastically from its data history during the online state estimation of

the system. Simulation results show that in presence of non-stationary Rayleigh

distributed noise the modified SIR filter gives better results compared to the other

considered filters. The new SIR PF also has been used to evaluate its performance

for a radar tracking system in presence of glint noise. Performance of new SIR PF

was found to be much better than SIR PF and MKF.

116

.
Chapter 8

Nonlinear State Estimation Using

Hybrid Robust Unscented Kalman
Filter

A robust unscented Kalman filter (UKF) and a hybrid robust unscented Kalman

filter have been proposed to estimate states of any nonlinear system in presence of

uncertainty. The robust UKF has been developed by combining a standard UKF

with an uncertainty estimator. The hybrid robust UKF has been designed using both

a UKF and a robust UKF, which adapts itself according to the value of normalized

innovation. The hybrid robust UKF appears to give better performance than the

conventional UKF and a newly proposed robust UKF both in absence and presence

of uncertainty, keeping a balance between estimation error and uncertainty. Two

examples are included to validate the performance of the proposed filter.

8.1 Introduction

The problem of dynamic state estimation is fundamental to control theory and signal

processing. The most widely used algorithm for state estimation of nonlinear systems

117
118

is the extended Kalman filter (EKF) [13, 14]. EKF introduces errors during lineariza-

tion compared to the unscented Kalman filter(UKF) [30] - [32]. Besides its higher

approximation accuracy, the latter avoids cumbersome evaluation of the Jacobian

and Hessian matrices making the algorithm easier to implement. For these advan-

tages UKF methods are sometimes employed instead of the standard EKF in wide

range of applications [112] - [115]. The best performance of any estimator (EKF or

UKF) is obtained when the process and measurement models along with their white

Gaussian noise statistics are known. However, in most physical systems, an approx-

imate model with parametric uncertainty and unknown external input is obtained.

Moreover, the process and measurement noise are colored and biased rather than

being zero mean and white. In recent years, such uncertainties in the plant model

is taken into account in several robust filtering techniques such as robust H∞ filter

[116] - [118], robust Kalman filter [119]-[121], robust minimum variance filters [122],

guaranteed cost filters [123], least square based robust filter [124], risk-sensitive filter

[125], smooth variable structure filter [126] etc. These differ in the kind of uncertainty

present in the plant. Unfortunately finding an effective and realistic model for the

plant uncertainty often poses much difficulty.

The present work addresses this problem. Here a robust UKF is developed which

requires no specific knowledge of the plant uncertainty model. The proposed modifi-

cation in UKF is said to be “robust” because it is capable of preserving filter perfor-

mance in spite of uncertainty and unknown disturbance in the plant description. This

uncertainty is also not classified a “norm bounded” one, as being commonly defined

in practice [116]-[130]. This filter is capable of estimating the states of any nonlinear

system in presence of various types of uncertainty - be it parameter uncertainty or

 119

unknown input or multiplicative noise or biased process noise etc, and that too in the

presence of large disturbances.

It is well known that the design of a robust filter is a trade off between robust-

ness and estimation accuracy. A robust filter designed for a graceful performance

degrades in presence of modeling errors (in literature [127]-[129]). So in the absence

of any unknown or time varying uncertainty the robust filter does not give desired

performance. Keeping this issue in mind, in this work a hybrid robust UKF has been

proposed which detects and adapts itself according to a change in the uncertainty in

the system. The hybrid robust filter consists of two filters- an UKF and a robust UKF

running parallely. At each time step it selects the estimated state of either of them,

depending on the uncertainty and based on the value of their respective normalized

innovation. So a hybrid robust UKF is expected to give favorable performance even

in absence of uncertainty.

The paper is organized as follows. In Section 8.2, the problem formulation is

described. Section 8.3, contains the estimation of uncertainty present in the plant

model using a low pass filter. The proposed robust UKF and hybrid robust UKF

has been described in Section 8.4 and Section 8.6 respectively. In Section 8.5, the

derivation of uncertainty estimation error covariance has been included. Simulation

results are shown in Section 8.7. Conclusions are drawn in Section 8.8.

8.2 Problem Formulation

Consider a stochastic nonlinear discrete time system.

xk+1 = f (xk , uk ) + wk (8.2.1)

120

where xk and uk are the stochastic state and known deterministic input respectively.

The process noise wk represents unknown input disturbances and any kind of un-

certainty which perturbs the system. The nonlinear measurement model is given

by

yk+1 = h(xk+1 ) + vk+1 (8.2.2)

vk is a zero mean white Gaussian noise with covariance Rk . Furthermore, wk and vk+1

are assumed to be uncorrelated. The problem is to estimate xk from a knowledge of

yk , corrupted by noise vk , and nominal system: xk+1 = f (xk , uk ).

8.3 Low Pass Filter for Uncertainty Estimation

In the state transition equation (8.2.1), the system noise wk denotes all kinds of uncer-

tainties which perturb the nominal system: xk+1 = f (xk , uk ). A quantity equivalent

to the uncertainty wk is expressed by weq,k , i.e.,

weq,k = xk+1 − f (xk , uk ) (8.3.1)

In time domain, this uncertainty in the current state can be estimated by

ŵk+1 = ŵeq,k (x̂k+1 ) = x̂k+1 − f (x̂k , uk ) (8.3.2)

where x̂k denotes estimate of x at k-th time step. A low pass filter is used below to

estimate the process uncertainty [130], excluding the high frequency noise,

ŵk+1 = Λŵeq,k (x̂k+1 ) = Λ(x̂k+1 − f (x̂k , uk )) (8.3.3)

where Λ is a diagonal matrix with low-pass-filter (LPF) entries. Although a lot of

candidates are possible for the selection of Λ matrix, here a first-order discrete filter
 121

has been considered: κi (z) = bi /(1 − ai z −1 )(i = 1, .., n) to simplify the problem.

Then, the matrix Λ will take the form

Λ(z) = diag(κ1 , κ2 , ...., κn ) (8.3.4)

By expanding (8.3.4), the state-space realization of (8.3.3) can be written as

     
ŵ1,k+1 b1 /(1 − a1 z −1 ) 0 ... 0 x1,k
     
 ŵ   0 b2 /(1 − a2 z −1 ) . . . 0   x 
 2,k+1     2,k 
 .. = .. .. ... .. + . 
 .   . . .   .. 
     
−1
ŵn,k+1 0 0 ··· bn /(1 − an z ) xn,k
    (8.3.5)
ŵ1,k+1 x1,k
   
 ŵ   x 
 2,k+1   2,k 
Where  ..  = ŵk , and  .  = (x̂k+1 − f (x̂k , uk )).
 .   .. 
   
ŵn,k+1 xn,k

Now from (8.3.5), we can write

ŵ1,k+1 (1 − a1 z −1 ) = b1 x1,k ŵ1,k+1 = ŵ1,k+1 a1 z −1 + b1 x1,k

ŵ2,k+1 (1 − a2 z −1 ) = b2 x2,k ŵ2,k+1 = ŵ2,k+1 a2 z −1 + b2 x2,k
.. ⇒ ..
. .
ŵn,k+1 (1 − an z −1 ) = bn xn,k ŵn,k+1 = ŵn,k+1 an z −1 + bn xn,k

Transforming the above equations in discrete time domain, we get

ŵ1,k+1 = a1 ŵ1,k + b1 x1,k

ŵ2,k+1 = a2 ŵ2,k + b2 x2,k
..
.
ŵn,k+1 = an ŵn,k + bn xn,k
122

Therefore,
       
ŵ1,k+1 a1 0 ... 0 ŵ1,k b1 0 ... 0 x1,k
       
 ŵ   0 a2 . . . 0   ŵ   0 b2 . . . 0  x 
 2,k+1     2,k     2,k 
 .. = .. .. . . ..  . + .. .. . . ..  .  (8.3.6)
 .   . . . .   ..   . . . .   .. 
       
ŵn,k+1 0 0 ··· an ŵn,k 0 0 ··· bn xn,k

ŵk+1 = F ŵk + G(x̂k+1 − f (x̂k , uk )) (8.3.7)

with F = diag(a1 , a2 , ..., an ), G = diag(b1 , b2 , ..., bn ) The discrete LPFs in (8.3.4)

should have unity DC gain, that is, κi (1) = 1(i = 1, ..., n). Thus, ai + bi = 1 and

F + G = I. Then, if the filter cut-off frequencies are decreased to zero (i.e. ai → 1

and bi → 0), the disturbance estimator will not work at all. It may be noted that wˆk

is the uncertainty estimate to be later included in the UKF algorithm.

8.4 Robust Unscented Kalman Filter

The robust UKF has been designed by including the uncertainty estimator (8.3.7)

in conventional UKF [30] - [32]. The algorithm given below is same as that of the

conventional UKF excepting the inclusion of estimated uncertainty ŵk in (8.4.8) and

uncertainty estimation error covariance Pw (k) in (8.4.10). The update equations for

ŵk and Pw (k) in (8.4.18) and (8.4.19) respectively, have been derived and included in

Section 8.5. In the following algorithm, the notation x̂k1 |k2 represents the estimate of

x at time k1 given observation up to, and including time k2

1. The n-dimensional random variable x(k) with mean x̂k|k and covariance Pk|k is
 123

approximated by 2n+1 weighted samples or sigma points selected by:

χo (k|k) = x̂k|k (8.4.1)

Wo = κ/(n + κ) (8.4.2)
√
χi (k|k) = x̂k|k + ( (n + κ)Pk|k )i (8.4.3)

Wi = 1/2(n + κ) (8.4.4)
√
χi+n (k|k) = x̂k|k − ( (n + κ)Pk|k )i (8.4.5)

Wi+n = 1/2(n + κ) (8.4.6)

√
( (n + κ)Pk|k )i is the i th row or column of the matrix square root of (n +

κ)Pk|k ),and Wi is the weight that is associated with i th point. The sample

covariance Pk|k is

2n
X
Pk|k = Wi (χi (k|k) − x̂k|k )(χi (k|k) − x̂k|k )T (8.4.7)
i=0

κ can be any number (positive or negative) provided that (n + κ) 6= 0. When

x(k) is assumed Gaussian, a useful heuristic is to select (n + κ) = 3 [31]. If

a different distribution is assumed for x(k) then a different choice of κ might

be more appropriate. Using the set of samples generated by (8.4.1)-(8.4.6), the

prediction procedure is as follows.

2. Each sigma point is instantiated through the process model (8.2.1) to yield a

set of transformed samples:

χi (k + 1|k) = f [χi (k|k), uk , k] + ŵk (8.4.8)

ŵk is the estimated uncertainty (8.3.7) included here.

124

3. The predicted mean is computed as

2n
X
x̂k+1|k = Wi χi (k + 1|k) (8.4.9)
i=0

4. The predicted covariance is computed as

2n
X
P (k+1|k) = Wi (χi (k+1|k)− x̂k+1|k )(χi (k+1|k)− x̂k+1|k )T +Pw (k) (8.4.10)
i=0

where Pw (k) is the uncertainty estimation error covariance. Derivation of Pw (k)

has been discussed in the Section 8.5.

5. Each predicted point is initiated through the observation model (8.2.2)

ζi (k + 1|k) = h[χi (k + 1|k), u(k)] (8.4.11)

6. The predicted observation is calculated by

2n
X
ẑk+1|k = Wi ζi (k + 1|k) (8.4.12)
i=0

7. The predicted innovation covariance is

2n
X
pzz (k + 1|k) = Rk+1 + Wi [ζi (k + 1|k) − ẑk+1|k ][ζi (k + 1|k) − ẑk+1|k ]T (8.4.13)
i=0

8. The predicted cross covariance is

2n
X
pxz (k + 1|k) = Wi [χi (k + 1|k) − x̂k+1|k ][ζi (k + 1|k) − ẑk+1|k ]T (8.4.14)
i=0

9. The UKF gain, updated state, and covariance respectively are

−1
Kk+1 = Pxz Pzz (8.4.15)

x̂k+1|k+1 = x̂k+1|k + Kk+1 (zk+1 − ẑk+1|k ) (8.4.16)

T
Pk+1|k+1 = Pk+1|k − Kk+1 Pzz Kk+1 (8.4.17)
 125

10. Uncertainty estimate is updated by

ŵk+1 = F ŵk + G(x̂k+1 − f (x̂k , uk )) (8.4.18)

11. Uncertainty estimation error covariance is updated (8.5.11) by (derived in Sec-

tion 8.5)

Pw (k + 1) = Pw (k) + (GKk+1 )R(GKk+1 )T + Qk (8.4.19)

8.5 Calculation of Uncertainty Estimation Error

Covariance

In this work instead of assuming wk to be a Gaussian white noise with fixed covariance,

it is estimated using the uncertainty estimator in (8.3.7). The covariance in turn can

be derived as follows.

From (8.4.1) and (8.4.8), we can write

χ0 (k + 1|k) = f [x̂k|k , u(k)] + ŵk (8.5.1)

Using (8.5.1) the estimated uncertainty (8.3.7) can be written as

ŵk+1 = F ŵk + G(x̂k+1 − {χ0 (k + 1|k) − ŵk }) (8.5.2)

= ŵk + G(x̂k+1 − {χ0 (k + 1|k)}) (8.5.3)

Since F + G = I.

Again, from (8.4.16),

x̂k+1|k+1 − x̂k+1|k = Kk+1 (zk+1 − ẑk+1|k ) (8.5.4)

Now, here χ0 (k + 1|k) is approximated by x̂k+1|k i.e.

χ0 (k + 1|k) ≈ x̂k+1|k (8.5.5)

126

So using (8.5.5) and (8.5.4), (8.5.3) can be written as

ŵk+1 = ŵk + GKk+1 (zk+1 − ẑk+1|k ) (8.5.6)

Now

wk+1 = wk + wk+1 − wk

= wk + ∆wk (8.5.7)

Where ∆wk is the difference in uncertainty between the kth and (k + 1)th instant

and is assumed to be white Gaussian with zero mean and covariance Qk

Now subtracting (8.5.6) from (8.5.7), we get

(wk+1 − ŵk+1 ) = (wk − ŵk ) − GKk+1 (zk+1 − ẑk+1|k ) + ∆wk (8.5.8)

w̃k+1 = w̃k − GKk+1 z̃k+1 + ∆wk (8.5.9)

Here (zk+1 − ẑk+1|k ) can be approximated by measurement noise vk which is uncorre-

lated with ∆wk and the uncertainty estimation error is w̃k+1 = (wk+1 − ŵk+1 ). Now

the uncertainty error covariance as per definition is

T
Pw (k + 1) = E[w̃k+1 w̃k+1 ] (8.5.10)

Finally, by using (8.5.9) and (8.5.10), the uncertainty estimation error covariance,

Pw (k + 1), can be derived as

Pw (k + 1) = Pw (k) + (GKk+1 )R(GKk+1 )T + Qk (8.5.11)

which is used in Section 8.4 in the algorithm of robust UKF.

 127

8.6 Hybrid Robust Unscented Kalman Filter

The robust filter is designed keeping a balance between uncertainty and estimation

error. So in absence of any unknown input or uncertainty, the performance of con-

ventional UKF is expected to be better compared to robust the UKF in terms of

estimation error. In this work a strategy is proposed to obtain desired performance

of the filters both in presence as well as absence of unknown inputs and uncertainty.

The UKF and robust UKF run in parallel. Whenever some unknown uncertainty is

detected, a suitable decision is taken to select either the estimated state of the UKF

or the robust UKF as the final state estimation.

The presence of uncertainty manifests itself in terms of a large innovation. A sim-

ple detection procedure for such an occurrence is based on the normalized squared

innovation as given below:

²ν (k) = ν(k)T S(k)ν(k) (8.6.1)

where, ν(k) and S(k) are the innovation and innovation covariance respectively, the

squared innovation ²ν (k) is calculated at each time step for both the filters, i.e., the

UKF and robust UKF and compared with each other. The estimated state of that

filter with a lower innovation is selected as the final estimated state. Instead of a

decision statistics based on a single sampling time, a moving average (or ‘moving

sum) of the normalized innovations squared over a sliding window of ‘s’ sampling

times is selected as:

k
X
²sν (k) = ²ν (j) (8.6.2)
j=k−s+1

Suppose at the kth instant, the UKF estimated mean is xukf (k|k) and covariance

is Pukf (k|k); the corresponding quantities for the robust UKF are xrukf (k|k) and
128

Prukf (k|k) respectively; the mean and covariance of ‘hybrid’ robust UKF are denoted

by xhrukf (k|k) and Phrukf (k|k) respectively. The output of this hybrid robust filter

is taken as the final estimated state of the system. The innovation of the UKF and

the robust UKF are denoted by ²sν,ukf (k) and ²sν,rukf (k) respectively. The decision

procedure is given as follows.

h i h i
• xukf (k + 1|k + 1), Pukf (k + 1|k + 1) = UKF xukf (k|k), Pukf (k|k)
h i h i
• xrukf (k + 1|k + 1), Prukf (k + 1|k + 1) = ROBUST UKF xrukf (k|k), Prukf (k|k)

• if ²sν,ukf (k + 1) > η²sν,rukf (k + 1)

h i h i
xhrukf (k + 1|k + 1), Phrukf (k + 1|k + 1) = xrukf (k + 1|k + 1), Prukf (k + 1|k + 1)
else
h i h i
xhrukf (k + 1|k + 1), Phrukf (k + 1|k + 1) = xukf (k + 1|k + 1), Pukf (k + 1|k + 1) ,
end

where η > 1, is a scaling parameter selected, depending on the admissible

uncertainty to the system. The value of η varies for different systems and un-

certainties.

8.7 Simulation Results

The robust UKF and hybrid robust UKF both are used to estimate the states of two

different uncertain nonlinear systems and the performance of the proposed filters have

been compared with the conventional UKF. In Example I, a target motion model with

unknown Ballistic coefficient and aerodynamic forces is taken from [131]. In Example

II, the Euler-discretized van der Pol oscillator is taken from [132]. The performances
 129

of all the three filters, i.e., the hybrid robust UKF, robust UKF and UKF have been

compared both in presence and in absence of unknown uncertainty.

8.7.1 Example I

A ballistic target motion model is considered with unknown ballistic coefficient [131].

In practical situations it is not realistic to assume a perfectly known ballistic coefficient

and aerodynamic forces during the entire flight.

Let k be a nonnegative integer number and ∆ be the time interval between two

consecutive radar measurements. The unknown augmented target state vector at

instant t = k∆ is defined by the five dimensional vector:

sk = [xk yk ẋk ẏk βk ]T (8.7.1)

where xk and yk are target positions, ẋk and ẏk target velocities along x and y axis

respectively, and βk is the unknown ballistic coefficient. It is assumed that βk evolves

in time according to a random walk model:

βk = βk−1 + wkβ (8.7.2)

where wkβ is a sequence of independent, identically distributed (IID) Gaussian vari-

ables with zero mean and variance q̃. The nonlinear ballistic motion model incorpo-

rating the random, time-varying ballistic coefficient is given by.

" #
0
sk+1 = Φsk + Bf (sk ) + B + wk (8.7.3)
−g
130

where g is the gravity acceleration (assumed constant), the matrices Φ and G are

given by
   
∆2
1 ∆ 0 0 0 0
   2 
 0 1 0 0 0   ∆ 0 
   
   
Φ= ∆ 0   0 ∆ 
2
 0 0 1  , B =  2 
(8.7.4)
   
 0 0 0 
1 0   0 ∆ 
  
0 0 0 0 1 0 0

and wk is a sequence of IID random vectors, with zero mean and nonsingular

covariance matrix  
q∆3 q∆2
0 0 0
 3 2 
 q∆2
q∆ 0 0 0 
 2 
 
Q=
 0 0 q∆3
3
q∆2
2
0 
 (8.7.5)
 q∆2

 0 0 q∆ 0 
 2 
0 0 0 0 q̃∆

where q is a positive real number and q̃ is the variance of wkβ in (8.7.2). Finally, the

nonlinear function f (sk ) in (8.7.3) denotes the aerodynamic force whose expression is
" #
q ẋk
g
f (sk ) = −0.5 ρ(yk ) ẋ2k + ẏk2 (8.7.6)
βk ẏk

where, ρ(·) is the air density which decays with altitude yk according to the exponen-

tial law:

ρ(y) = c1 exp(−c2 y) (8.7.7)

where, c1 = 1.227, c2 = 1.093 × 10−4 for y < 9144 m, and c1 = 1.754, c2 = 1.49 × 10−4

for y ≥ 9144 m, assuming that the measurements are generated by a conventional

tracking radar.The two dimensional observation vector at instant t = k∆ is zk =

[rk εk ]T , where rk and εk are the measured range and elevation angle respectively.
 131

The measurement equation is

" p #
x2k + yk2
zk = + vk (8.7.8)
arctan( xykk )
vk is modeled as an Gaussian random sequence, also independent of initial state s0

and wk with zero mean and a nonsingular covariance matrix.

à !
σr2 0
R= (8.7.9)
0 σε2

The state model (8.7.3) and measurement model (8.7.8) have been used to simulate

the trajectory with parameters g = 9.8m/s2 , ∆ = 2s, q = q̃ = 5, σr = 150 m, and

σε = 0.017 rad. The initial state s0 is taken to be a Gaussian random vector with mean

m0 = [232000 m 2290 cos(190◦ ) m/s 88000 m 2290 sin(190◦ ) m/s 40000 kg.m−1 .s−2 ]T

and diagonal covariance matrix Σ0 with Σ0 (1, 1) = Σ0 (3, 3) = 10002 m2 , Σ0 (2, 2) =Σ0 (4, 4) =

502 m2 .s−2 and Σ0 (5, 5) = 25002 , kg2 .m−2 .s−4 . To evaluate the performance of the

proposed filter, it is assumed that the ballistic coefficient and aerodynamics forces

are completely unknown, replacing f (·) by zero in (8.7.3). The target trajectory is

tracked over 50 time steps using the UKF, robust UKF and hybrid robust UKF with

unknown ballistic coefficient and aerodynamics forces i.e. putting f (·) = 0 in the

prior update stage of the filter. For the robust UKF the parameters are selected as
h iT
G = 0.2I5×5 , F = 0.8I5×5 , initial mean of uncertainty ŵ0 = 0 0 0 0 0 , and

initial uncertainty covariance Pw (0) = Q. In case of the hybrid robust UKF, the scal-

ing parameter η = 2.5 and sliding window of sampling times s = 5 has been selected.

For all the three filters the corresponding root mean square error (RMSE) curves

for estimated target position, in the coordinates x and y using all the three filters

are shown in Figs (8.1-8.2) assuming unknown ballistic coefficient and aerodynamics

forces. While those are shown in Figs (8.3-8.4), taking known ballistic coefficient and
132

aerodynamics forces. A total of 100 Monte Carlo simulation runs were used to com-

pute the RMSE plots. The curves show that in presence of uncertainty Fig (8.1-8.2)

the UKF diverges compared to the robust UKF. Also in absence of any uncertainty

the hybrid UKF follows the conventional UKF Figs (8.3-8.4).

rms Position error along x−axis(metre)

3500
UKF
Robust UKF
3000 Hybrid Robust UKF

2500
RMSE(m)

2000

1500

1000

500

0
0 10 20 30 40 50
iteration number

Figure 8.1: RMS error along x-axis with unknown ballistic coefficient and aerody-
namic forces
 133

rms Position error along y−axis(metre)

7000
UKF
Robust UKF
6000 Hybrid Robust UKF

5000
RMSE(m)

4000

3000

2000

1000

0
0 10 20 30 40 50
iteration number

Figure 8.2: RMS error along y-axis with unknown ballistic coefficient and aerody-
namic forces

rms Position error along x−axis(metre)

1000
UKF
900 Robust UKF
Hybrid Robust UKF
800

700
RMSE(m)

600

500

400

300

200

100
0 10 20 30 40 50
iteration number

Figure 8.3: RMS error along x-axis with known ballistic coefficient and aerodynamic
forces
134

rms Position error along y−axis(metre)

1400
UKF
Robust UKF
1200 Hybrid Robust UKF

1000
RMSE(m)

800

600

400

200

0
0 10 20 30 40 50
iteration number

Figure 8.4: RMS error along y-axis with known ballistic coefficient and aerodynamic
forces

8.7.2 Example II

The Euler-discretized van der Pol oscillator is given by [132]

" # " # " #
x1,k x1,k−1 + T x2,k−1 0
= + uk−1 (8.7.10)
x2,k −T x1,k−1 + (T + 1 − T x21,k−1 )x2,k−1 1

where xk = [x1,k x2,k ]T is the state vector, uk−1 is the external input assumed to be

unknown and T is the sampling interval. The measurement equation is given by

zk = [1 1]xk + vk (8.7.11)

The measurement noise is zero mean white Gaussian with covariance R. For simula-

tion we set x0 = [1 1]T , T = 0.1, R = 0.04 and

T sin(2kT ), if kT < 10s or kT ≥ 30s

uk−1 = { T sin(2kT ) + 0.5, if 10s ≤ kT < 20s (8.7.12)
T sin(2kT ) − 0.5, if 20s ≤ kT < 30s
 135

The performance of all the filters (i.e. UKF, robust UKF, and hybrid robust UKF)

have been compared assuming the input uk−1 to be unknown and known separately.

Whenever uk−1 is assumed to be unknown, uk−1 = 0 is set in the priori update stage

of the filter. For all the filters the initial mean and covariance matrices are set to
h iT
x̂(0|0) = 0.5 1.5 and P (0|0) = 0.5I2×2 and for the known input, the process

noise "covariance
# is selected
" as Q#= 10−6 I2×2 . The robust UKF is implemented with
0 0 0 0 h iT
G= ,F = , initial value of mean uncertainty w0 = 0 0
0 0.2 0 0.8
and initial uncertainty covariance Pw (0|0) = Q. The scaling parameter η = 1.5 and

sliding window of sampling times s = 5 are selected for the hybrid robust UKF. The

performance of the three filters have been compared regarding RMSE of x2,k over 100

Monte Carlo simulation runs. Three distinct cases have been taken into account.

(i) considering unknown input (Fig 8.5)

(ii) considering known input (Fig 8.6)

(iii) considering unknown input, but with a larger value of process noise covariance
" #
10−6 0
Q only in case of UKF (Fig 8.7). In this case Q = is set for UKF.
0 .05

Simulation results shows that in case 1, the robust UKF and hybrid robust UKF

give better results than UKF (Fig 8.5). In case 2 and case 3, comparable results

are obtained using the UKF and robust UKF but hybrid robust UKF gives the best

result among the filters in these two cases (Fig 8.6-8.7).

136

RMS estimation error of x2

7
UKF
Robust UKF
6 Hybrid Robust UKF

4
RMSE

0
0 50 100 150 200 250 300 350 400
iteration number

Figure 8.5: RMS estimation error of x2 with unknown input

RMS estimation error of x2

1.4
UKF
Robust UKF
1.2 Hybrid Robust UKF

0.8
RMSE

0.6

0.4

0.2

0
0 50 100 150 200 250 300 350 400
iteration number

Figure 8.6: RMS estimation error of x2 with known input

 137

RMS estimation error of x2

1.4
UKF
Robust UKF
1.2 Hybrid Robust UKF

0.8
RMSE

0.6

0.4

0.2

0
0 50 100 150 200 250 300 350 400
iteration number

Figure 8.7: RMS estimation error of x2 with unknown input with a larger value of
Qk in case of UKF

8.8 Conclusion

This work proposes the design of a hybrid robust UKF to estimate the states of any

nonlinear model in presence of unknown, unmodeled uncertainty. The robust UKF is

a combination of an uncertainty estimator with a standard UKF and can be considered

to be a useful filtering technique which produces reliable estimates in presence of large

modeling errors, unknown inputs, heavy process noise and unknown disturbances.

The proposed hybrid robust UKF preserves a balance between estimation error and

uncertainty. It detects the presence of uncertainty and adapts the system accordingly.

Two examples have been taken to compare the performances of the proposed filters.

Example I, considers a ballistic target tracking problem with an unknown ballistic

coefficient and aerodynamic forces in filter design. It is seen that the robust UKF

and hybrid robust UKF both give much better results than the UKF. The filters are
138

once more tested with known ballistic coefficient and aerodynamic forces (absence of

uncertainty), where it is seen that the estimation error of robust UKF is little higher

than UKF and hybrid robust UKF. The simulation results show that the hybrid

robust UKF gives best performance and that too in presence as well as absence of

unknown uncertainty. In Example II, a van der Pol oscillator has been considered.

The simulation results have been compared among the three filters (i.e. UKF, robust

UKF and hybrid robust UKF) considering cases with a known and unknown input

separately. The proposed filters are once more compared with the case in which the

UKF is employed with a larger process noise covariance matrix (where the input is

unknown). Simulation results show that in case of an unknown input the robust UKF

gives better results than the UKF and as well as the UKF with a higher process noise

covariance. The hybrid robust UKF is seen to give favorable results both with known

and unknown input. So it can be said the hybrid robust UKF delivers acceptable

results for any nonlinear system in presence or absence of uncertainty.

Chapter 9

Conclusion

The work embodied in this thesis attempts to identify areas in dynamic state esti-

mation in stochastic systems where there is scope for further improvement either in

the design of the models or the algorithms. The thesis mainly deals with dynamic

state estimation of uncertain system from noisy measurements. The design and de-

velopment of some estimators which builds upon previously successful approaches but

improves upon them in different aspects have been documented in this thesis. The

estimators in this thesis have been developed for different forms of uncertainty in the

systems / systems models, and also in the measurements/ measurement models to

achieve their best possible performance.

The following highlights the overall contributions of this thesis:

• For a linear system in presence of zero mean white Gaussian process noise and

measurement noise, a recursive state estimator has been derived from a basic

‘maximum a posteriori’ parameter estimator, named as recursive maximum a

posteriori estimator (RMAPE) and it has also been shown that the well known

139
140

Kalman filter (KF) algorithm (which is a MMSE estimator) can be mathemat-

ically derived from it - i.e. this work also shows an alternative way to derive

the KF.

• The problem of state estimation of ballistic target missiles from noisy RF seeker

measurements has also been dealt with. Comparing the performances of EKF,

UKF and SIR PF regarding estimation error and computational load, the EKF

is found to give better performance among all the filters under consideration.

The measurements obtained from RF seeker contain uncertainties in form of

lag, spike, data lose, jumps and discontinuity. These uncertainties in measure-

ments, other than zero mean white Gaussian noise, contribute degradation in

filter performance. Some modifications in EKF algorithm and system model

is incorporated to process those uncertainties for real-time state estimation of

ballistic target missiles.

• Significant efforts are expended to develop filter algorithms for state estimation

assuming zero mean white Gaussian measurement noise- but this is not always

true in real world situations. Thus, furthermore, in this thesis, the state estima-

tion in presence of non-stationary non-Gaussian measurement noise has been

considered. A generalized algorithm is proposed to estimate the parameters of

the PDF of a noise signal, which may be a) stationary or non-stationary, b)

Gaussian or non-Gaussian, c) zero mean or non-zero mean and d) uni-modal

or multi-modal. The proposed generalized algorithm of parameter estimation

of the noise sequence improves the filter performance in presence of any type

of white measurement-noise, by incorporating the proposed algorithm into the

likelihood function of the SIR particle filter algorithm.

 141

• Finally, after considering the uncertainty in measurements, we have concen-

trated on the various forms of uncertainty in the state space model. A robust

hybrid unscented Kalman filter is proposed for estimating the states of non-

linear systems in the presence or absence of any inadequately modeled and/or

undefined uncertainty, e.g. parameter uncertainties, modeling errors, unknown

inputs, heavy process noise and also unknown disturbances. In this filter, there

is no need of any prior knowledge of the uncertainties in the system model.

The problem of dynamic state estimation in the presence of uncertainty is a broad

area with existing and evolving challenges. As the field progresses, esp. in part under

the requirements of new engineering applications primarily in aerospace and military

technologies, different approaches, including hybrid ones will be tried and tested. In

this thesis we have primarily tried to improve the performance of nonlinear filters in

presence of different types of uncertainties in system and measurement model.

9.1 Future Scope of the Work

It is an evident fact that finding of an efficient optimal nonlinear filter is still a open

challenge research problem. The work embodied in this thesis provides improved in-

sights into the modeling and design of suboptimal nonlinear filters. This work also

creates new information based on which some progress in the design of optimal non-

linear optimal filters can be made in the future. A very important area within this

is the problem of real time estimation and areas in mission critical applications with

high degree of uncertainty but zero margin of error. Another area where not enough

attention have been given for the testing and adaptation of the existing theory of
142

dynamic state estimation in presence of uncertainty in the area of biomedical engi-

neering which poses new types of engineering challenges esp. new forms of uncertainty

and noise not encountered in other domains of engineering. One important attempt

should be the integration of all existing estimation algorithms in an integrated soft-

ware framework for direct and easy selection, combination and use to cater for the

complete domain of problems requiring state estimation.

Appendix A

Extended Kalman Filter

The EKF [13]-[16] is a set of mathematical equations which uses an underlying pro-

cess model to make an estimate of the current state of a system and then corrects

the estimate using any available sensor measurements. Using this predictor-corrector

mechanism, it approximates an optimal estimate due to the linearization of the pro-

cess and measurement models. The algorithm of the EKF is presented in the following

omitting theoretical details.

Nonlinear Dynamic model:

xk = fk−1 (xk−1 ) + wk−1 (A.0.1)

Where, xk is the state vector, wk−1 is the zero mean white Gaussian noise with

covariance Qk−1 .

Nonlinear measurement model:

zk = hk (xk ) + vk (A.0.2)

Where, zk is the sensor measurement, vk is the zero mean white Gaussian noise with

covariance Rk .

The nonlinear implementation equations are given as follows:

143
144

• The predicted a-priori state estimate:

x̂k|k−1 = fk−1 (x̂k−1|k−1 ) (A.0.3)

• The predicted measurement:

ẑk|k−1 = hk (x̂k−1|k−1 ) (A.0.4)

• 1 st order Linear approximation equations (Jacobian Matrix):

[1] ∂fk
φk−1 = |x=x̂k−1|k−1 (A.0.5)
∂x
[1] ∂hk
Hk = |x=x̂k|k−1 (A.0.6)
∂x

• The a priori covariance

[1] [1]
Pk|k−1 = φk−1 Pk−1|k−1 φk−1 + Qk−1 (A.0.7)

• gain matrix
[1]T [1]T
Kk = Pk|k−1 Hk (Hk Pk|k−1 Hk + Rk )−1 (A.0.8)

• a posteriori mean
[1]
x̂k|k = x̂k|k−1 + Kk [zk − Hk x̂k|k−1 ] (A.0.9)

• a posteriori covariance

[1]
Pk|k = [I − Kk Hk ]Pk|k−1 (A.0.10)

The EKF gives good approximation for slightly nonlinear system in presence of Gaus-

sian noise. But for highly nonlinear the performance of the EKF degrades because of

error due to linearization. The UKF has been shown to give better performance than

a standard EKF since it better approximates the nonlinearity.

Appendix B

Unscented Kalman Filter

The basic promise behind the UKF [30]-[33] is that it is easier to approximate a

Gaussian distribution than to approximate any arbitrary nonlinear function. Instead

of using Jacobian matrices for the purpose of linearization, it uses the true nonlinear

model involving the state and/or measurement equations while approximating the

probability density function of the state vector. This density is specified by a set of

deterministically chosen samples or sigma points whose sample mean and covariances

are x̂k−1|k−1 and Pk−1|k−1 respectively. The nonlinear function is applied to each

of these points in turn to yield a transformed sample, and the predicted mean and

covariance are calculated from these transformed points.

• The n-dimensional random variable xk−1 with mean x̂k−1|k−1 and covariance

Pk−1|k−1 is approximated by 2n+1 weighted samples or sigma points selected

145
146

by the algorithm. (Only algorithm of UKF is briefly described in this section)

χ0 (k − 1|k − 1) = x̂k−1|k−1 (B.0.1)

W0 = κ/(n + κ) (B.0.2)
√
χi (k − 1|k − 1) = x̂k−1|k−1 + ( (n + κ)Pk−1|k−1 )i (B.0.3)

Wi = 1/2(n + κ) (B.0.4)
√
χi+n (k − 1|k − 1) = x̂k−1|k−1 − ( (n + κ)Pk−1|k−1 )i (B.0.5)

Wi+n = 1/2(n + κ) (B.0.6)

√
( (n + κ)Pk−1|k−1 )i is the i-th row or column of the matrix square root of

(n + κ)Pk−1|k−1 ),and Wi is the weight that is associated with i-th point. The

sample covariance Pk−1|k−1 is

2n
X
Pk−1|k−1 = Wi (χi (k−1|k−1)− x̂k−1|k−1 )(χi (k−1|k−1)− x̂k−1|k−1 )T (B.0.7)
i=0

κ can be any number (positive or negative) providing that (n + κ) 6= 0

The set of samples generated by (B.0.1)-(B.0.6), the prediction procedure is as

follows.

• Each sigma points is instantiated through the process model (A.0.1) to yield a

set of transformed samples

χi (k|k − 1) = f [χi (k − 1|k − 1)] (B.0.8)

• The predicted mean computed as

2n
X
x̂k|k−1 = Wi χi (k|k − 1) (B.0.9)
i=0
 147

• The predicted covariance computed as

2n
X
Pk|k−1 = Wi (χi (k|k − 1) − x̂k|k−1 )(χi (k|k − 1) − x̂k|k−1 )T (B.0.10)
i=0

• Initiate each predicted points through the observation model

ζi (k|k − 1) = h[χi (k|k − 1)] (B.0.11)

• The prediction observation is calculated by

2n
X
ẑk|k−1 = Wi ζi (k|k − 1) (B.0.12)
i=0

• The prediction innovation covariance

2n
X
pzz (k|k−1) = Rk + Wi [ζi (k|k−1)− ẑ(k|k−1)][ζi (k|k−1)− ẑk|k−1 ]T (B.0.13)
i=0

• The prediction cross covariance

2n
X
pxz (k|k − 1) = Wi [χi (k|k − 1) − x̂k|k−1 ][ζi (k|k − 1) − ẑk|k−1 ]T (B.0.14)
i=0

• The UKF gain, Updated state, and covariance are respectively

−1
Kk = Pxz Pzz (B.0.15)

x̂k|k = x̂k|k−1 + Kk (zk − ẑk|k−1 ) (B.0.16)

Pk|k = Pk|k−1 − Kk Pzz KkT (B.0.17)

For some highly nonlinear system UKF shows better performance than standard

EKF as it better approximate nonlinearity. But both EKF and UKF works on the

assumption of Gaussian PDF. But the performance of these filter are not well when

the true density is non-Gaussian specially for multi-modal or heavily skewed PDF. In

that case The particle filter yield an improvement in performance compared to EKF

and UKF.
148

.
Appendix C

Particle Filter

The PF [17]-[20] is a computer-based method for implementing an optimal recursive

Bayesian filter by Monte Carlo simulations. Instead of analytic solution or numerical

approximation of a given nonlinear/non-Gaussian problem, it performs a considerable

amount of computations in order to approximate the posterior PDF. The central idea

is to represent the required PDF by a set of random samples (particles). As the

number of particle is increased, the representation of the required PDF becomes

more accurate. The PF in general requires the knowledge of

1. The initial state PDF p(x0 = pN (x0 ; x̂0|0 , P0|0 )), where x̂0|0 and P0|0 are respec-

tively mean and covariance of the initial state.

2. The likelihood function p(zk |xk ) = pN (zk ; h(xk ), Rk ))

3. The statistics of process noise wk . The PDF of process noise is p(wk ) =

pN (wk ; 0; Qk ).

The PF algorithm can be described as follows. Assume that we have a set of

random samples (particles) Σk = {xk (i) : i = 1, .., n} from the posterior density

at time k, i.e. p(xk |Zk ). The PF is an algorithm for propagating and updating

149
150

the set of random samples Σk to a new set of random samples at time k + 1 ,

Σk+1 = {xk+1 (i) : i = 1, .., n}, which are approximately distributed as the posterior

density p(xk+1 |Zk+1 ). There are different particle filter algorithms such as sampling

importance resampling (SIR) filter, auxiliary sampling importance resampling (ASIR)

filter, regularized particle filter (RPF) [17, 18]. In this work we use the sequential im-

portance sampling (SIR) or the bootstrap algorithm. The bootstrap filter propagates

the random sample set by the following steps.

• Prediction: Each sample in Σk−1 is passed through the state equation A.0.1 to

obtain samples from the prior density at time step k

x∗k (i) = fk−1 (xk−1 (i)) + wk−1 (i) (C.0.1)

where wk is a sample drawn from the PDF of process noise p(wk ) .

• Update: i) Calculate the likelihood function p(zk |x∗k (i)) for each sample in the

set Σ∗k = x∗k (i) : i = 1, ..., n . ii) Calculate the normalized weights for each

samples
p(zk |x∗k (i))
qk (i) = n (C.0.2)
X
p(zk |x∗k (j))
j=1

The weights qk (i) represent the probability mass associated with element i of

Σ∗k .

• Resample n times from the discrete distribution defined by Σ∗k and qk = {xk (i) :

i = 1, ..., n} to generate samples Σk = {x∗k (i) : i = 1, ..., n} so that for any i0 ,

P r{xk (i0 ) = x∗k (i)} = qk (i)

 151

The MMSE estimate of xk is computed as the mean of particles in Σk−1 :

n
1X
x̂k|k = xk (i) (C.0.3)
n i=1

The choice of the number of particles n is very important in the PF design. Some

recent PF [20] schemes can operate with a smaller number of particles without a sig-

nificant loss in estimation accuracy. PF is considered to be most optimal filter in case

of nonlinear non-Gaussian system compared to other filters. But the computational

load is much higher compared to others.

152

.
Bibliography

[1] H. J. Kushner, “Dynamical equations for optimal non-linear filtering,” J. Differ-

ential Equations, vol. 3, pp. 179-190, 1967.

[2] A. H. Jazwinski, Estimation with Applications to Tracking and Navigation, New

York, Academic, 1970.

[3] Masaaki Kijima, Markov processes for stochastic modeling, CRC Press, 1997.

[4] H. W. Sorenson, “On the develpment of practical nonlinear filters,” Information

Science, vol. 7, pp. 253-270, 1974.

[5] H. Tanizaki, Nonlinear Filters: Estimation and Applications, New York:

Springer-Verlag, 2nd ed., 1996.

[6] B. D. O. Anderson and J. B. Moore, Optimal Filtering, Englewood Cliffs, NJ:

Printice-Hall, 1979.

[7] S. C. Kramer and H. W. Sorenson, “Recursive Bayesian estimation using piece-

wish constant approximations,” Autometica, vol. 24, no. 6, pp. 789-801, 1988.

[8] K. Kastella, “Finite difference methods for nonlinear filtering and automatic target
recognition” in Multitarget-Multisensor Tracking (Y. Bar-Shalom and W. D. Blair,
eds.), vol. III, ch. 5, Norwood, MA: Artech House, 2000.

153
154

[9] L. D. Stone, “A Bayesian approach to multiple-target tracking” in Handbook of

Multisensor Data Fusion (D. L. Hall and J. Linas, eds.), ch. 10, Boca Raton, FL:
CRC Press, 2001.

[10] F. Martineric and P. Forster, “Data association and tracking using hidden
Markov models and dynamic programming,” in Proc. Conf. ICASSP 92, 1992.

[11] L. R. Rabiner, “A tutorial on hidden Markov models and selected applications

in speech recognition,” in Proc. of the IEEE, vol. 77, pp. 257-285, February 1989.

[12] H. W. Sorenson and D. L. Alspach, “Recursive Bayesian estimation using Gaus-

sian sum approximations,” Autometica, vol. 7, pp. 465-448, 1971.

[13] Robert Grover Brown and Patric Y. C. Hwang, Introduction to Random Signals
and Applied Kalman Filtering, John Wiley and Sons, 3rd Edition, 1996.

[14] Mohinders S. Grewal and Angus P. Andrews, Kalman Filtering Theory and Prac-
tice, Prentice Hall, Englewood Cliffs, New Jersey, 1993

[15] Yaakov Bar-Shalom, X. Rong Li and Thiagalingam Kirubaranjan, Estimation

with Applications to Tracking and Navigation, John Wiley and Sons, 2001.

[16] P.S.Maybeck, Stochastic Models, Estimation and Control, New York, Academic,
1982, vol. 2.

[17] Branko Ristic, Sanjeev Arulampalam and NeilGordon, Beyond the Kalman Filter
Particle Filters for Tracking Applications, Artech House., Boston, London, 2004.

[18] M. Sanjeev Arulampalam, Simon Maskell, Neil Gordon and Tim Clapp, “A Tu-
torial on Particle Filters for Online Nonlinear/Non-Gaussian Bayesian Tracking,”
IEEE Transactions on Signal Processing, vol. 50, no. 2, Feb 2002.
 155

[19] A. Doucet, S. J. Godsill and C. Andrieu, “On sequential simulation-based meth-

ods for Bayesian fltering,” Statistics and Computing , vol. 10, no. 3, pp. 197-208,
2000.

[20] A. Doucet, N. De Freitas and N. Gordon, Sequential Monte Carlo Methods in

Practice, Springer Verlag, 2001.

[21] N. J. Gordon, D. J. Salmond and A.F.M. Smith, “Novel approach to

nonlinear/non-Gaussian Bayesian state estimation,” IEE Proceedings, vol. 140,
no. 2, April 1993.

[22] W. Gilks, S. Richardson and D. Spiegelhalter, Markov Chain Monte Carlo in

practice, Chapman and Hall, 1996.

[23] M. Pitt and N. Shephard, “Filtering via simulation: Auxilary particle filters,”
Journal of the American Statistical Association, vol. 94, no. 446, pp. 590-599,
1999.

[24] C. Musso, N. Oudjane, and F. Le Gland. “Improving regularized particle filters,”

in Sequential Monte Carlo Methods in Practice, Statistics for Engineering and
Information Science(A. Doucet, N. de Freitas, and N. Gordon), Springer–Verlag,
New York, chapter 12, pp. 247–271, 2001.

[25] W. R. Gilks and C. Berzuini, “Following a moving target - Monte Carlo inference
for dynamic Baysian models,” Jouranal of Royal Statistical Society, B, vol. 63, pp.
127-146, 2001.

[26] Arnaud Doucet and Simon Godsill and Christophe Andrieu, ”On sequential
Monte Carlo sampling methods for Bayesian filtering,” Statistics And Comput-
ing, vol. 10, no. 3, pp. 197 - 208. 2000.
156

[27] R. van der Merwe and A. Doucet and N. De Freitas and E. Wan, ”The Unscented
Particle Filter,” Tech. Rep CUED/F-INFENG/TR 380, Cambridge University
Engineering Department, 2000.

[28] S. McGinnity and G.W. Irwin, ”Multiple model bootstrap filter for maneuvering
target tracking,” IEEE Transactions on Aerospace and Electronic Systems, vol.
36, no. 3, pp. 1006 - 1012. 2000.

[29] D. S. Angelova, Tz. A. Semerdjiev, V. P. Jilkov and E. A. Semerdjiev, ”Ap-

plication of a Monte Carlo method for tracking maneuvering target in clutter,”
Mathematics and Computers in Simulation, vol. 55, pp. 15 - 23. 2001.

[30] S.J. Julier, J.K. Uhlmann and H.F. Durrant-Whyte, “A new approach for filter-
ing nonlinear system,” Proc. Am. Contr. Conf., Seattle, WA, 1995, pp. 1628-1632.

[31] S.J. Julier and J.K. Uhlmann, “A new extension to Kalman filter to nonlin-
ear systems,” Proc. Aerosense:11th Int. Symp. Aerosp/Defense Sensing, Simulat.
contr., Orlando, FL, 1997.

[32] S.J. Julier, J.K. Uhlmann and H.F. Durrant-Whyte, “A new method for the
nonlinear transformation of means and covariances in filters and estimators ,”
IEEE Transaction on Automatic Control, vol. 45, no. 3, March 2000.

[33] S.J. Julier and J.K. Uhlmann, “Unscented Filtering and Nonlinear Estimation,”
Proceedings of IEEE, vol. 92, no. 3, pp. 401-422, March 2004.

[34] Bowling C. and Jones R., “Motion Compensated Image Coding with a Com-
bined Maximum A Posteriori and Regression Algorithm,” IEEE Transaction on
communication, vol 33 , no. 8, pp. 844 - 857, Aug 1985.

[35] Dembo A. and Zeitouni O., “Maximum a posteriori estimation of time-varying

ARMA processes from noisy observations,” IEEE Transaction on Signal Process-
ing, vol. 36 , no. 4, pp. 471 - 476, April 1988.
 157

[36] Hero A.O., “Timing estimation for a filtered Poisson process in Gaussian noise,”
IEEE Transaction on Information Theory, vol 37 , no. 1, pp. 92 - 106, Jan 1991.

[37] Fessler J.A., Erdogan H. and Wei Biao Wu, “Exact distribution of edge-
preserving MAP estimators for linear signal models with Gaussian measurement
noise,” IEEE Transaction on Image Processing, vol 9 , no. 6, pp. 1049 - 1055,
June 2000.

[38] Johnston L.A. and Krishnamurthy V., “An improvement to the interacting mul-
tiple model (IMM) algorithm,” IEEE Transaction on Signal Processing, vol 49 ,
no. 12, pp. 2909 - 2923, Dec 2001.

[39] JoonBeom Kim, Stuber G.L. and Ye Li , “Low-complexity iterative channel

estimation for serially concatenated systems over frequency-nonselective Rayleigh
fading channels,” IEEE Transaction on Wireless Communications, vol. 6 , no. 2,
pp. 438 - 442, Feb 2007.

[40] Hua Fu and Pooi Yuen Kam, “MAP/ML Estimation of the Frequency and Phase
of a Single Sinusoid in Noise,” IEEE Transaction on Signal Processing, vol. 55,
no. 3, pp. 834-845, March 2007.

[41] Schizas I.D., Giannakis G.B., Roumeliotis S.I. and Ribeiro A., “Consensus in
Ad Hoc WSNs With Noisy LinksPart II: Distributed Estimation and Smoothing
of Random Signals,” IEEE Transaction on Signal Processing, vol 56 , no. 4, pp.
1650 - 1666, April 2008.

[42] Jie Gao and Huaping Liu, “Low-Complexity Map Channel Estimation for Mobile
MIMO-OFDM Systems,” IEEE Transaction on Wireless Communications, vol. 7
, no. 3, pp. 774 - 780, March 2008.
158

[43] Alessio Benavoli, Luigi Chisci and Alfonoso Farina, “Estimation of Constrained
Parameters with Guaranteed MSE Improvement,” IEEE Transaction on Signal
Processing, vol. 55, no. 4, April 2007.

[44] B. Ristic, A. Farina, D. Benvenuti and M.S. Arulampalam, “Performance bounds

and comparison of nonlinear filters for tracking a ballistic object on re-entry”, IEE
Proc.-Radar Sonar Navig., vol. 150, no. 2, pp. 65-70, April 2003.

[45] A. Farina, B. Ristic and D. Benvenuti, “Tracking a Ballistic Target: Comparison

of Several Nonlinear Filters”, IEEE Transactions on Aerospace and Electronic
Systems, vol. 38, no. 3, pp. 854-867, July 2002.

[46] A. Farina, B. Ristic and D. Benvenuti, “Improved sequential Monte Carlo filter-
ing for ballistic target tracking”, IEEE Transactions on Aerospace and Electronic
Systems, vol. 41, no. 3, pp. 1103- 1108, July 2005.

[47] Jacques Waldmann, “Tracking Filters and Models for Seeker Applications”,
IEEE Transactions on Aerospace And Electronic Systems, vol. 37, no. 3, pp. 965-
977, July 2001.

[48] Jacques Waldmann, “Line-of-Sight Rate Estimation and Linearizing Control of

an Imaging Seeker in a Tactical Missile Guided by Proportional Navigation”,
IEEE Transactions on Control Systems Technology, vol. 10, no. 4, pp. 556- 567,
July 2002.

[49] Ananthasayanam M. R., A. K. Sarkar, A. Bhattacharya, P. K. Tiwari and

Prashant Vohra, “Nonlinear Observer State Estimation From Seeker Measure-
ments and Seeker-Radar Measurements Fusion”, American Institute of Aeronau-
tics and Astronautics AIAA Guidance, Navigation, and Control Conference and
15 - 18 August 2005, San Francisco, California

[50] George M. Siouris, Aerospace Avionics System, Academic Press, 1993

 159

[51] Hilton R.D., Martin D.A. and Blair W.D., “Tracking with time-delayed data in
multi sensor systems.”, NSWCDD/TR-93/351, Dahlgren, VA, Aug.1993.

[52] Blackman S.S. and Popoli R., “Design and Analysis of Modern Tracking Sys-
tems”, Boston: Artech House,1999.

[53] Bar-Shalom Y., “Update with out-of-sequence measurements in tracking: Exact

solution”, IEEE Transaction on Signal Processing, vol. 38, no. 3, pp. 769-778,
July 2002.

[54] Nettleton E.W. and Durrant-Whyte H., “Delayed and asequent data in decentral-
ized sensing networks.”, In Proceedings of SPIE Conference, vol. 4571, Oct/Nov
2001.

[55] Zhang K. S., Li X. R. and Zhu Y. M., “Optimal update with out-of-sequence
observations for distributed filtering.”, In Proceedings of the Fifth International
Conference on Information Fusion,Annapolis, MD, July 2002.

[56] S. Vathsal, A. K. Sarkar, Ranjit Das and D. Ghose,“Realistic Pursuer Evader En-
gagement Simulation With PN and Bang-Bang Guidance”, Proceedings of AIAA
Conference on Guidance, Navigation, and Control 15 - 18 August 2005, San Fran-
cisco, California.

[57] Sadhu S. and Ghoshal T. K., “Observer Based Estimation Approaches for Seeker
Filtering,” The Institution of Electrical Engineering, Savoy place, London WC2R
OBL, UK, 2001.

[58] P. Garnell and D.J. East, Guided Weapon Control Systems, pergamon press
Elmsford, N. Y., 1977.

[59] Z Wang, DWC Ho and X Liu, “Variance-constrained filtering for uncertain

stochastic systems with missing measurements”, IEEE Transaction on Automatic
Control, vol. 48, no. 7, pp. 1254- 1258, July 2003.
160

[60] Z Wang, F Yang, DWC Ho and X Liu, “Robust finite-horizon filtering for stochas-
tic systems with missing measurements”, IEEE Signal Processing Letter, vol. 12,
no. 6, pp. 437- 440, June 2005.

[61] Z Wang, F Yang, DWC Ho and X Liu, “Robust H∞ filtering for stochastic
time-delay systems with missing measurements”, IEEE Transaction on Signal
Processing, vol. 54, no. 7, pp. 2579 - 2587, July 2006.

[62] B Sinopoli, L Schenato, M Franceschetti, K. Poolla, M.I. Jordan and S.S. Sastry,
“Kalman filtering with intermittent observations”, IEEE Transactions on Auto-
matic Control, vol. 49, no. 9, pp. 1453- 1464, Sept 2004.

[63] Wu Weng-Rong, “Target tracking with glint noise,” IEEE Trans. Aerosp. Elec-
tron. Syst, vol. 29, no. 1, pp. 174-185, Jan. 1993.

[64] B. Borden and M. Mumford, “A statistical glint/radar cross section target

model,” IEEE Trans. Aerosp. Electron. Syst., vol. 19, no. 1, pp. 781-785, Sept.
1983.

[65] G. Hewer, R. Martin and J. Zeh, “Robust preprocessing for Kalman filtering of
glint noise,” IEEE Trans. Aerosp. Electron. Syst., vol. 23, pp. 120-128, 1987.

[66] C. J. Masreliez, “Approximate non-Gaussian filtering with linear state and ob-
servation relations,” IEEE Transactions on Automatic Control, vol. 20, no. 2, pp.
107110, 1975.

[67] C. J. Masreliez and R. D. Martin, “Robust Bayesian estimation for the lin-
ear model and robustifying the Kalman filter,” IEEE Transactions on Automatic
Control, vol. 22, no. 3, pp. 361371, 1977.

[68] W. R. Wu and A. Kundu, “Kalman filtering in non-Gaussian environment using

efficient score function approximation.” In Proceedings of 1989 IEEE International
Symposium on Circuits and Systems, 1989, pp. 413416.
 161

[69] Wu, W. R. and P. P. Cheng, “A nonlinear IMM algorithm for maneuvering target
tracking,” IEEE Transactions on Aerospace and Electronic Systems, vol. 30, no.
3, pp. 875885, 1994.

[70] E. Daeipour and Y. Bar-Shalom, “An interacting multiple model approach for
target tracking with glint noise,” IEEE Transactions on Aerospace and Electronic
Systems, vol. 31, no. 2, pp. 706715, 1995.

[71] Wen-Hong Zhu and T. Lamarche, “Velocity estimation by using position and
acceleration sensors,” IEEE Trans. Ind. Electron., vol. 54, no. 5, pp. 2706-2715,
Oct. 2007.

[72] W. Panusittikorn, Min Cheol Lee and P.I. Ro, “Modeling and sliding-mode
control of friction-based object transport using two-mode ultrasonic excitation,”
IEEE Trans. Ind. Electron., vol. 51, no. 4, pp. 917 - 926, Aug. 2004.

[73] Seiichiro Katsura, Yuichi Matsumoto and Kouhei Ohnishi, “Modeling of force
sensing and validation of disturbance observer for force control,” IEEE Trans.
Ind. Electron., vol. 54, no. 1, pp. 530 - 538, Feb. 2007.

[74] Yaakov Bar-Shalom and X. Rong Li and Thiagalingam Kirubaranjan, Estimation

with Applications to Tracking and Navigation, John Wiley and Sons, 2001.

[75] Y.S. Suh, “Attitude estimation by multiple-mode kalman filters,” IEEE Trans.
Ind. Electron., vol. 53, no. 4, pp. 1386- 1389, Aug. 2006.

[76] U.A. Abeyratne, A.P. Petropulu and J.M. Reid, “Higher order spectra based
deconvolution of ultrasound images,” IEEE Trans. Acoust. Speech Signal Process.,
vol. 36, no. 5, pp. 642-650, 1998.

[77] S.R. Kadaba, S. B. Gelfand and R. L. Kashyap, “Recursive estimation of images

using non-Gaussian autoregressive models,” IEEE Trans. Image. Process., vol. 7,
no. 10, pp. 1439- 1452, 1998.
162

[78] M.S. Rizk, W. Zgallai, P. Hardiman and J.O. Riordan, “Blind homomorphic
analysis of abnormalities in ECG signals,” Proceedings on the IEE Colloquium on
Blind Deconvolution-Algorithm and Applications, pp. 195-199, 1995.

[79] S. Fang, W. LI and D. Zhu, “Modeling and simulation of non-Gaussian correlated

clutter,” Proceedings of CIE International Conference on Radar, pp. 195-199,
1996.

[80] J.M. Mendel, “Tutorial on higher-order statistics (spectra) in signal processing

and system theory: Theory, results and some applications,” Proc. of the IEEE,
vol. 79, no. 3, pp. 278-305, March 1991.

[81] R.O. Duda and P.E. Hart, Pattern Classification and Scene Analysis, New York:
John Wiley and Sons, 1973.

[82] R.A. Redner and H.F. Walker, “Mixture densities maximum likelihood and the
EM algorithms,” SIAM Review, vol. 26, no. 2, April 1984.

[83] B.S. Everitt and D.J. Hand, Finite mixture Distributions, London, U.K.: Chap-
man and Hall, 1981.

[84] G.J. Mclachlan and K.E. Basford, Mixture Models, New York: Marcel Dekker,
1988.

[85] D.M. Titterington, A.F.M Smith and U.E. Makov, Statistical Analysis of Finite
Mixture Distributions, Chichester, U.K.: Wiley, 1985.

[86] S.M. Verbout, J. M. Ooi, J. T. Ludwing and A. Oppenheim,“Parameter esti-

mation for autoregressive Gaussian-mixture processes: The EMAX algorithm,”
IEEE Trans. Signal Process., vol. 46, no. 10, pp. 2744 - 2756, 1998

[87] D. Sengupta and S.M. Kay, “Efficient Estimation of Parameters Non-Gaussian

Autoregressive Processes,” IEEE Trans. Acoust. Speech. Signal Process., vol. 37,
no. 6, pp. 785 - 794, Jun. 1989.
 163

[88] Y. Zhao, X. Zhuang and S.J. Ting, “Gaussian mixture density modeling of non-
Gaussian source for autoregressive process,” IEEE Trans. Signal Process., vol. 43,
no. 4, pp. 894 - 903, April 1995.

[89] Hong-Zhou Tan and T.W.S. Chow, “Blind and total identification of arma mode
in higher order cumulants domain,” IEEE Trans. Ind. Electron., vol. 46, no. 6,
pp. 1233 - 1240, Dec. 1999.

[90] E.A. Ibatoulline, “Parameter estimation of non-Gaussian probability density of

one-dimensional signals and interferences by iterative method of maximum like-
lihood,” IEEE International Symposium on Electromagnetic Compatibility, EMC
’03, vol. 1, pp. 363 - 366, May 2003.

[91] E.A. Ibatoulline, “Iterative method of maximun likelihood for estimation of non-
Gaussian probability density of signals and interferences,” IEEE International
Symposium on Electromagnetic Compatibility, pp. 524 - 528, May 2002.

[92] Weifeng Liu, P.P. Pokharel and J.C. Principe, “Correntropy: Properties and
applications in non-Gaussian signal processing,” IEEE Trans. Signal Process.,
vol. 55, no. 11, pp. 5286 - 5298, Nov. 2007.

[93] Youshen Xia and M.S. Kamel, “A generalized least absolute deviation method
for parameter estimation of autoregressive signals,” IEEE Trans. Neural Netw.,
vol. 19, no. 1, pp. 107 - 118, Jan. 2008.

[94] Arpita Mukherjee and Aparajita Sengupta, “Parameter estimation of a signal

alongwith non-stationary non-Gaussian noise,” The 33rd Annual Conference of
the IEEE Industrial Electronics Society (IECON), pp. 2429-2433, Nov. 2007.

[95] Zuo Dongguang, Han Chongzhao, Bian Shutan Zhenglin and Zhu Hongyan,
“Tracking of maneuvering target in glint noise environment,” Proceedings of the
164

Sixth IEEE International Conference of Information Fusion, vol. 2, pp. 1394 -

1399, 2003.

[96] WU Wen-Rong and WU Kuo-Guan, “Adaptive identification of non-

Gaussian/non-stationary glint noise,” IEICE Transactions on Fundamentals of
Electronics, Communications and Computer Sciences, vol. E82-A, no. 12, pp.
2783-2792, 1999.

[97] R. A. Redner and H.F. Walker, “On the convergence properties of the EM algo-
rithm,” Ann. Statist., vol. 11, no. 1, pp. 95-103, 1983.

[98] Lichtenauer J., Reinders M. and Hendriks E., “Influence of the observation like-
lihood function on particle filtering performance in tracking applications,” Pro-
ceedings of Sixth IEEE International Conference on Automatic Face and Gesture
Recognition, pp. 767-772, 17-19 May 2004

[99] Daniel L. Alspach and Harold W. Sorenson, “Nonlinear Bayesian Estimation

Using Gaussian Sum Approximations ,” IEEE Transactions on Autouatic Control,
vol. 17, no. 4, pp. 439 - 448, Aug 1972.

[100] Rong Chen and Jun S. Liu, “Mixture Kalman Filters ,” Journal of Royal Sta-
tistical Society, vol. 62, part 3, pp. 493 - 508, 2000.

[101] A. Doucet, N.J. Gordon, V. Krishnamurthy, “Particle filters for state estimation
of jump Markov linear systems,” IEEE Transactions on Signal Processing, vol. 49,
no. 3, pp. 613-624, Mar. 2001.

[102] Thomas Schn and Fredrik Gustafsson, “Marginalized Particle Filters for Mixed
Linear/Nonlinear State-Space Models,” IEEE Transactions on Signal Processing,
vol. 53, no. 7, July 2005.
 165

[103] Jitendra K. Tugnait and Abraham H. Haddad, “Adaptive Estimation in Lin-

ear Systems with Unknown Markovian Noise Statistics ,” IEEE Transactions on
Information Theory, vol. 22, no. 1, pp. 66 - 78, Jan 1980.

[104] Jayesh H. Kotecha and Petar M. Djuric, “Gaussian Sum Particle Filtering,”
IEEE Transactions on Signal Processing, vol. 51, no. 10, pp. 2602 - 2612, Oct.
2003.

[105] J. H. Kotecha and P.M. Djuric, “Gaussian particle filtering,” IEEE Transactions
on Signal Processing, vol. 51, pp. 25932602, Oct. 2003.

[106] Arpita Mukherjee, Dr. Aparajita Sengupta, “Parameter Estimation of a Signal

Alongwith Non-Stationary Non-Gaussian Noise,” The 33rd Annual coonference of
the IEEE Industrial Electronics Society(IECON), pp. 2429-2433, Nov 5-8, 2007.

[107] X.Wang and R. Chen, “Adaptive Bayesian multiuser detection for synchronous
CDMA with Gaussian and impulsive noise,” IEEE Trans. Signal Process., vol. 47,
no. 7, pp. 20132028, Jul. 2000.

[108] W. R. Wu, “Maximum likelihood identification of glint noise,” IEEE Trans.

Aerosp. Electron. Syst., vol. 32, pp. 4151, Jan. 1996.

[109] Rong Chen, Xiaodong Wang and Jun S. Liu, “Adaptive Joint Detection and
Decoding in Flat-Fading Channels via Mixture Kalman Filtering,” IEEE Trans-
actions on Information Theory, vol. 46, no. 6, pp. 2079 - 2094, Sept 2000.

[110] W. Wu and P. Cheng, “Nonlinear IMM algorithm for maneuvering target track-
ing,” IEEE Trans. Aerosp. Electron. Syst., vol. 30, no. 3, pp. 875-884, July 1994.

[111] X. Li and V. Jilkov, “Survey of maneuvering target tracking. Part V: Multiple-

model methods,” IEEE Trans. Aerosp. Electron. Syst., vol. 41, no. 4, 2005.
166

[112] C. Vijayakumar, and R. Rajagopal, “Passive target tracking by unscented fil-

ters,” In Proceedings of the IEEE International Conference on Industrial Tech-
nology, pp. 129133, Jan. 2000.

[113] X. Yi, and L. Li, “ Single observer bearings-only tracking with the unscented
Kalman filter,” In Proceedings of the IEEE International Conference on Commu-
nications, Circuits and Systems, vol. 2, pp. 901905, June 2004.

[114] R. van der Merwe, E. A. Wan, and S. J. Julier,“Sigma-point Kalman filters for
nonlinear estimation and sensor-fusion: Applications to integrated navigation”, In
Proceedings of AIAA Guidance Navigation and Controls Conference, Providence,
RI, 2004.

[115] R. Zhan, J. Wan, “Iterated Unscented Kalman Filter for Passive Target Track-
ing” Aerospace and Electronic Systems, IEEE Transactions on vol 43, no. 3, pp.
1155 - 1163, July 2007.

[116] Jaewon Seo, Myeong-Jong Yu, Chan Gook Park, And Jang Gyu Lee,“An Ex-
tended Robust H∞ Filter for Nonlinear Constrained Uncertain Systems”, IEEE
Transactions on Signal Processing, vol.54, no.11, November 2006.

[117] Shengyuan Xu, James Lam, and Xuerong Mao“Delay-Dependent H∞ Control

and Filtering For Uncertain Markovian Jump Systems With Time-Varying De-
lays,”IEEE Transactions on Circuits And Systems, vol. 54, no.9, September 2007.

[118] Carlos E. de Souza, Karina A. Barbosa, and Alexandre Trofino Neto “Robust
H∞ Filtering for Discrete-Time Linear Systems With Uncertain Time-Varying
Parameters,”IEEE Transactions on Signal Processing, vol. 54, no.6, June 2006.

[119] Fuwen Yang, Zidong Wang, and Y. S. Hung,“Robust Kalman Filtering for Dis-
crete Time-Varying Uncertain Systems With Multiplicative Noises,” IEEE Trans-
actions on Automatic Control, vol. 47, no. 7, July 2002.
 167

[120] Zhe Dong and Zheng You “Finite-Horizon Robust Kalman Filtering For
Uncertain Discrete Time-Varying Systems With Uncertain-Covariance White
Noises”,IEEE Signal Processing Letters, vol. 13, no. 8, August 2006.

[121] K. Xiong, H. Zhang, L. Liu, “Adaptive robust extended Kalman filter for non-
linear stochastic systems,” Control Theory and Applications, IET, vol. 2, no. 3,
pp. 239 - 250, March 2008.

[122] U. Shaked, C. E. de Souza, “Robust minimum variance filtering,” IEEE Trans-

actions on Signal Processing, vol. 43, no. 11, pp. 2474-2483, Nov 1995.

[123] T. Li, L. Guo, “Optimal Fault-Detection Filtering for Non-Gaussian Systems

via Output PDFs,” IEEE Transactions on Systems, Man and Cybernetics, Part
A, vol. 39, no. 2, pp. 476-481, March 2009.

[124] W. S. Ra, I. H. Whang, J. Y. Ahn and J. B. Park,“Recursive robust least squares

estimator for time-varying linear systems with a noise Corrupted measurement
matrix,” IET Control Theory Appl., vol. 1, no. 1, January 2007.

[125] S. Dey, J. B. Moore, “Risk-sensitive filtering and smoothing via reference prob-
ability methods,” IEEE Transactions on Automatic Control, vol. 42, no. 11, pp.
1587-1591, Nov 1997

[126] S. Habibi, “The Smooth Variable Structure Filter,” Proceedings of the IEEE ,
vol. 95, no. 5, pp. 1026-1059, May 2007 .

[127] U. Shaked and C. E. de Souza, Robust Estimation and Faliure Detection: A

Concise Treatment, London, U.K.: Springer, 1998.

[128] H. Li and M. Fu, “A Linear Matrix Inequility Approach to Robust H∞ filtering,”

IEEE Trans. Signal Process., vol. 45, pp. 23382350, Sept 1997.
168

[129] Lihua Xie, Yeng Chai Soh and Carlos E. de Souza, “Robust Kalman Filtering
for Uncertain Discrete-Time Systems,” IEEE Transactions on Automatic Control,
vol. 39, no. 6, June 1994

[130] S. J. Kwon, “Robust Kalman filtering with perturbation estimation process for
uncertain systems,” IEE Proc.-Control Theory Appl., vol. 153, no. 5, September
2006.

[131] M. G. S. Bruno, A. Pavlov, “Improved Sequential Monte Carlo Filtering for Bal-
listic Target Tracking,” IEEE Transactions on Aerospace and Electronic Systems,
vol. 41, no. 3, pp. 1103-1108, July 1999.

[132] B. Teixeira, J. Chandrasekar, H. J. Palanthandalam-Madapusi, L. Torres, L. A.

Aguirre, and D. S. Bernstein, “Gain-Constraint Kalman Filtering for Linear and
Nonlinear Systems,” IEEE Transactions on Signal Processing, vol. 56, no. 9, pp.
4113-4123, Sept. 2008.

[133] Arpita Mukherjee, Aparajita Sengupta, “State Estimation of Linear Stochas-

tic System Using Recursive MAP Estimator”, Proceedings of World Academy of
Science, Engineering and Technology, Volume 32, August 2008. (Accepted)

[134] Arpita Mukherjee, Aparajita Sengupta, “State Estimation of Linear Stochas-

tic System Using Recursive Maximum a Posteriori Estimator”, Asian Journal of
Control. (revised paper submitted)

[135] Arpita Mukherjee, Subhankar Bhowmik, Aparajita Sengupta, “State Estima-

tion of a Ballistic Missile in presence of Bad Seeker Measurement and Model
Uncertainty”, JOURNAL OF SYSTEMS SCIENCE AND ENGINEERING, Vol
17, No1, Page no 23-29, June 2008.
 169

[136] Arpita Mukherjee, Aparajita Sengupta, “Kinematic State Estimation of a Bal-

listic Target Missile from Bad Seeker Measurements”, IEEE International confer-
ence on Avionics Systems ICAS-2008, Feb 2008, pp 246-246.

[137] Arpita Mukherjee, Aparajita Sengupta, “State Estimation of Ballistic Target

Missile from Seeker Measurements using Nonlinear Filters”, National System con-
ference (NSC-2007), Dec 14-15, 2007.

[138] Arpita Mukherjee, Aparajita Sengupta, “Seeker Filter Design of Ballistic Tar-
get Missile Incorporating Seeker Dynamics”, International Research Publication
House - International Journal of Industrial Electronics and Control (Accepted)

[139] Arpita Mukherjee, Aparajita Sengupta, “Parameter Estimation of a Signal

along with Non-Stationary Non-Gaussian Noise”, 33rd Annual Conference of the
IEEE Industrial Electronics Society IEEE IECON 2007, pp 2429-2433, Nov 5-8,
2007.

[140] Arpita Mukherjee, Aparajita Sengupta, “Estimating the Probability Density

Function of a Non-Stationary Non-Gaussian Noise”, IEEE Transaction of Indus-
trial Electronics, (Revised twice, final decision awaited)

[141] Arpita Mukherjee, Aparajita Sengupta, “Likelihood Function Modeling of Par-

ticle Filter in Presence of Non-Stationary Non-Gaussian Measurement Noise”,
Signal Processing (revised paper submitted)

[142] Arpita Mukherjee, Aparajita Sengupta, “Likelihood Function Modeling of Par-

ticle Filter in Presence of Non-Stationary Non-Gaussian Measurement Noise”, The
2009 International Conference on Graphic and Image Processing (ICGIP 2009)
(Accepted)