Introduction and Background

Least Squares Estimation

The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Fundamentals of Kalman Filtering and

Estimation in Aerospace Engineering

Christopher DSouza
[email protected]

NASA / Johnson Space Center

Houston, Texas

20 March 2013

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 1 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Outline

Introduction and Background

Concepts from Probability Theory
Linear and Nonlinear Systems
Least Squares Estimation
The Kalman Filter
Stochastic Processes
The Kalman Filter Revealed
Implementation Considerations and Advanced Topics
The Extended Kalman Filter
Practical Considerations
Advanced Topics
Conclusions

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 2 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Introduction and Background
Concepts from Probability Theory
Linear and Nonlinear Systems

Least Squares Estimation

The Kalman Filter

Stochastic Processes
The Kalman Filter Revealed

Implementation Considerations and Advanced Topics

The Extended Kalman Filter
Practical Considerations
Advanced Topics

Conclusions
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 3 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Why Estimate?

We estimate without even being conscious of it

Anytime you walk down the hallway, you are estimating, your
eyes and ears are the sensors and your brain is the computer
In its essence, estimation is nothing more than taking noisy
sensor data, filtering the noise, and producing the best state
of the vehicle

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 3 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

What Do We
Estimate?

As NASA engineers, we estimate a variety of things

Position, Velocity, Attitude
Mass
Temperature
Sensor parameters (biases)
These quantities are usually referred to as the states of the
system
We use a variety of sensors to accomplish this task
Inertial Measurement Units (IMUs)
GPS Receivers (GPSRs)
LIDARs
Cameras
These sensors are used to determine the states of the

system

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 4 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

A Brief History
of Estimation

Estimation has its origins in the work of Gauss and his

innovation called Least Squares Estimation
He was interested in computing the orbits of asteroids and
comets given a set of observations
Much of the work through WWI centered around extensions to
Least Squares Estimation
In the interval between WWI and WWII, a number of
revolutionary contributions were made to sampling and
estimation theory
Norbert Weiner and the Weiner Filter
Claude Shannon and Sampling Theory
Much of the work in the first half of the Twentieth Century
focused on analog circuitry and the frequency domain

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 5 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Modern Estimation
and Rudolf Kalman

Everything changed with the confluence of two events:

The Cold War and the Space Race
The Advent of the Digital Computer and Semiconductors
A new paradigm was introduced: State Space Analysis
Linear Systems and Modern Control Theory
Estimation Theory
Optimization Theory
Rudolf Kalman proposes a new approach to linear systems
Controllability and Observability

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 6 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Rudolf Kalman
and His Filter


In 1960 Kalman wrote a paper in an obscure ASME journal
entitled A New Approach to Linear Filtering and Prediction
Problems which might have died on the vine, except:
In 1961, Stanley Schmidt of NASA Ames read the paper and
invited Kalman to give a seminar at Ames
Schmidt recognized the importance of this new theory and
applied it to the problem of on-board navigation of a lunar
vehicle after all this was the beginning of Apollo
This became known as the Kalman Filter
Kalmans paper was rather obtuse in its nomenclature and
mathematics
It took Schmidts exposition to show that this filter could be
easily mechanized and applied to a real problem
The Kalman Filter became the basis for the on-board

navigation filter on the Apollo CSM and LM
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 7 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Types of Estimation

There are basically two types of estimation: batch and

sequential
Batch Estimation
When sets of measurements taken over a period of time are
batched and processed together to estimate the state of a
vehicle at a given epoch
This is usually the case in a ground navigation processor
Sequential Estimation
When measurements are processed as they are taken and the
state of the vehicle is updated as the measurements are
processed
This is done in an on-board navigation system

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 8 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Types of Sensors

Inertial Measurement Units (IMUs)

GPS Recievers
Magnetometers
Optical Sensors
Visible Cameras
IR Cameras
LIDARs (Scanning and Flash)
RF sensors
Radars (S-band and C-band)
Range and Range-rate from Comm
Altimeters
Doppler Velocimeters

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 9 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Introduction and Background
Concepts from Probability Theory
Linear and Nonlinear Systems

Least Squares Estimation

The Kalman Filter

Stochastic Processes
The Kalman Filter Revealed

Implementation Considerations and Advanced Topics

The Extended Kalman Filter
Practical Considerations
Advanced Topics

Conclusions
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 10 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Why do we care about this probability stuff?

Information: the negative reciprocal value of probability .

Claude Shannon

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 10 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Concepts from
Probability Theory


A random variable is one whose value is subject to
variations due to chance (randomness) it does not have a
fixed value; it can be discrete or continuous
A coin toss: can be heads or tails discrete
The lifetime of a light bulb continuous
A probability density function (pdf), p (x ), represents the
likelihood that x occurs
Always non-negative
Satisfies
Z
p () d = 1

The expectation operator, E [f (x )], is defined as

Z
E [f (x )] = f () p () d

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 11 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Concepts from
Probability Theory Mean and Variance


The mean (or first moment) of a random variable x, denoted
by x, is defined as
Z

x = E [x ] = p () d

h i
The mean-square of a random variable x, E x 2 , is defined
as
Z
h i
E x2 = 2 p () d

The variance (or second moment) of a random variable x,
denoted by 2x , is
Z
h 2
i
2x = E [x E (x )] = ( E ())2 p () d

h i
= E x 2 x2
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 12 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Concepts from
Probability Theory Mean and Variance

of a Vector
, is defined as
The mean of a random nvector x, x
Z Z Z

= E [x]
x = p () d

The (co-)variance of random nvector x, Px , is defined as

Z iT
h T
i h ih
Px ] [x x
= E [x x ] = p () d

x1 x1 x2 x1 xn
2



2x2 x2 xn
x1 x2
= . .. .. ..

.. . . .


x1 xn x2 xn 2xn


The covariance is geometrically represented by an error

ellipsoid.
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 13 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Concepts from
Probability Theory The Gaussian

Distribution
The Gaussian probability distribution function, also called
the Normal distribution1 or a bell curve, is at the heart of
Kalman filtering
We assume that our random variables have Gaussian pdfs
We like to work with Gaussians because they are completely
characterized by their mean and covariance
Linear combinations of Gaussians are Gaussian
The Gaussian distribution of random nvector x, with a mean
and covariance Px , is defined as
of x
1 x)T P
(x 1
x (x
x)
pg (x) = e 2
(2)n/2 |Px |
1
Physicist G. Lippman is reported to have said, Everyone believes in the

normal approximation, the experimenters because they think it is a mathematical
theorem, the mathematicians because they think it is an experimental fact.
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 14 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Concepts from
Probability Theory The Gaussian

Distribution

We can show that

x)T P 1
Z
1 (x x (x
x)
e 2 dx = 1
Rn (2)n/2 |Px |
If a random process is generated
by a sum of other (non-Gaussian)
random processes, then, in the
limit, the combined distribution
approaches a Gaussian distribution
(The Central Limit Theorem)

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 15 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Introduction and Background
Concepts from Probability Theory
Linear and Nonlinear Systems

Least Squares Estimation

The Kalman Filter

Stochastic Processes
The Kalman Filter Revealed

Implementation Considerations and Advanced Topics

The Extended Kalman Filter
Practical Considerations
Advanced Topics

Conclusions
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 16 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Linear Systems

A system is a mapping from input signals to output signals,

written as: w (t ) = L (v (t ))
A system is linear if for all input signals v (t ), v1 (t ), and v2 (t )
and for all scalars ,
L is additive: L (v1 (t ) + v2 (t )) = L (v1 (t )) + L (v2 (t ))
L is homogeneous: L (v (t )) = L (v (t ))
For a system to be linear, if 0 is an input, then 0 is an output:
L (0) = L (0 v (t )) = 0 L (v (t )) = 0
If the system does not satisfy the above two properties, it is
said to be nonlinear
If L (v (t )) = v (t ) + 1, is this linear?
It is not because for v (t ) = 0, L (0) = 1 , 0

Lesson: Some systems may look linear but they are not!

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 16 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Nonlinear Systems
and the Linearization Process

Despite the beauty associated with linear systems, the fact of

the matter is that we live in a nonlinear world.
So, what do we do? We make these nonlinear systems into
linear systems by linearizing
This is predicated on a Taylor series approximation which we
deploy as follows: Given a nonlinear system of the form:
X = f(X, t ), we linearize about (or expand about) a nominal
trajectory, X? (with X? = f(X? , t )), as

f
!
X (t ) = f (X? , t ) + (X X? ) +
X X=X?

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 17 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Nonlinear Systems
and the State Transition Matrix


If we let x(t ) = X X? and let F(t ) = Xf
 
, then we get
X=X?

x = F(t )x with x(t0 ) = x0

The solution of this equation is
Rt
F() d
x(t ) = e t0
x0 = (t , t0 )x0

where (t , t0 ) is the State Transition Matrix (STM) which

satisfies
t , t0 ) = F(t )(t , t0 ) with (t0 , t0 ) = I
(
The STM can be approximated (for F(t ) = F = a constant) as
Rt
F() d 1
(t , t0 ) = e t0
= e F(t t0 ) = I + F(t t0 ) + F2 (t t0 )2 +
2
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 18 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

A Bit More About the State Transition Matrix

The State Transition Matrix (STM) is at the heart of practical
Kalman filtering. In its essence it is defined as
X(t )
(t , t0 ) =
X(t0 )
As the name implies, it is used to transition or move perturbations
of the state of a nonlinear system from one epoch to another, i.e.
X(t ) 
X(t ) X? (t ) = X(t0 ) X? (t0 )
  
x(t ) = (t , t0 )x(t0 )
X(t0 )
2
In practical Kalman filtering, we use a first-order  approximation
f(X, t ) 
(t , t0 ) I + F(t0 ) (t t0 ) = I + (t t0 )
X X=X0


2
In cases of fast dynamics, we can approximate the STM to second-order as:
(t , t0 ) I + F(t0 ) (t t0 ) + 21 F2 (t0 ) (t t0 )2
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 19 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Introduction and Background
Concepts from Probability Theory
Linear and Nonlinear Systems

Least Squares Estimation

The Kalman Filter

Stochastic Processes
The Kalman Filter Revealed

Implementation Considerations and Advanced Topics

The Extended Kalman Filter
Practical Considerations
Advanced Topics

Conclusions
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 20 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

How do we Implement This?

Never do a calculation unless you already know the answer.

John Archibald Wheelers First Moral Principle

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 20 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Context of Least Squares Estimation

Least Squares estimation has been a mainstay of engineering
and science since Gauss invented it to track Ceres circa 1794
It has been used extensively for spacecraft state estimation,
particularly in ground-based navigation systems
The Apollo program had an extensive ground station network
(MSFN/STDN) coupled with sophisticated ground-based
batch processors for tracking the CSM and LM
A set of measurements (or several sets of measurements)
taken over many minutes and over several passes from
different ground stations would be batched together to get a
spacecraft state at a particular epoch
Least Squares estimation is predicated on finding a solution
which minimizes the square of the errors of the model

y = Hx + 

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 21 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Least Squares

Problem


The problem is as follows: given a set of observations, y, subject to
measurement errors ( ), find the best solution, x
, which minimizes
the errors, i.e.
1 T 1
min J =   = (y Hx)T (y Hx)
2 2
To do this we take the first derivative of J with respect to x and set
it equal to zero as

J 1
" #
= T
(y Hx) (y Hx) )T H = 0
= (y Hx
x x 2
x=x

, is
Therefore, the optimal solution, x
  1
= HT H
x HT y

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 22 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Weighted
Least Squares (WLS) Problem


Suppose now we are given measurements y, whose error has a
measurement covariance of R. How can we get the best estimate,
which minmizes the errors weighted by the accuracy of the
x
measurement error (R1 )? The problem can be posed as
1 T 1 1
min J = R  = (y Hx)T R1 (y Hx)
2 2
Once again, we take the first derivative of J with respect to x and
set it equal to zero as
J 1
" #
= T 1
(y Hx) R (y Hx) )T R1 H = 0
= (y Hx
x x 2 x=
x

, is
Therefore, the optimal solution, x
  1
= HT R1 H
x HT R1 y

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 23 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The WLS Problem

with A Priori Information


Suppose we need to find the best estimate of the state, given
measurements y, with measurement error covariance R, but we
with covariance
are also given an a priori estimate of the state, x

P. This problem can be posed as
1 1
min J = (y Hx)T R1 (y Hx) + (x
x)T P
1 ( x
x)
2 2
As before, we take the first derivative of J with respect to x and set
it equal to zero as
J 

)T R1 H (x
= (y Hx )T P
x 1 = 0
x x=x


, is
Therefore, the optimal solution, x
  1 h i
x 1
= HT R1 H + P 1 x
HT R1 y + P

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 24 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Nonlinear Batch
Estimation

In general, the system of interest will be nonlinear of the form

Yk = h(Xk , tk ) +  k

How do we get the best estimate of the state X? Well, first we

linearize about a nominal state X?
k
(with xk = Xk X?
k
) as

h 

h(X? h(X? Xk X?
 
Yk = + xk , tk ) +  k = k) + + + k
k
X Xk =X?k k


k = h 
Defining H X Xk =X?
 we get the following equation
k

k xk +  k
yk = H

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 25 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Nonlinear Batch
Estimation at an Epoch


In batch estimation, we are interested in estimating a state at an
epoch, say X0 , with measurements taken after that epoch say, at
tk . How can we obtain this? Well, we use the state transition matrix
as follows
Xk X? ?
 
k = (tk , t0 ) Xk Xk xk = (tk , t0 )x0
so that we can map the measurements back to the epoch of
interest as
k (tk , t0 )x0 +  k = Hk x0 +  k
yk = H
The least squares solution (over all the p measurements) is
p 1 p
0 X?
X T 1 X
1 H R yi + P
0 = Hi Ri Hi + P
x


T 1 x
1 = X
0 i i 0 0 0
i =1 i =1

This is called the normal equation.

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 26 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Nonlinear
Batch Estimation Algorithm

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 27 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Batch Filter Example Aircraft Tracking

Given a ground station tracking an

airplane, moving in a straight line at
a constant speed, with only bearing
measurements, we are interested in
knowing the speed of the airplane
and its position at the beginning of
the tracking pass (x0 , y0 , u0 , v0 ). The
equations are

x (t ) = u0 (t t0 ) + x0
y (t ) = v0 (t t0 ) + y0
" #
1 y (t )
(t ) = tan
x ( t )

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 28 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Batch Filter Example Aircraft Tracking (II)

The initial guess is

The measurements are:
985
105
X0 =
1.5

10

with initial covariance

100 0 0 0
0 100 0 0
P0 =
0 0 1 0

0 0 0 1

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 29 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Batch Filter Example Aircraft Tracking (III)

After 7 iterations the following results are obtained:

Parameter Truth Initial Guess Converged State

x0 1000 985 983.5336
y0 100 105 99.3470
u0 -3 -1.5 -2.9564
v0 12 10 11.7763

Lesson: The x component is not readily observable. But that is

not surprising since angles do not provide information along the
line-of-sight.

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 30 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Something to remember

One must watch the convergence of a numerical code as carefully

as a father watching his four year old play near a busy road.
J. P. Boyd

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 31 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Introduction and Background
Concepts from Probability Theory
Linear and Nonlinear Systems

Least Squares Estimation

The Kalman Filter

Stochastic Processes
The Kalman Filter Revealed

Implementation Considerations and Advanced Topics

The Extended Kalman Filter
Practical Considerations
Advanced Topics

Conclusions
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 32 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Need for

Careful Preparation

Six months in the lab can save you a day in the library

Albert Migliori, quoted by J. Maynard

in Physics Today 49, 27 (1996)

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 32 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Stochastic Processes
The Linear First-Order

Differential Equation
Let us look at a first-order differential equation for x (t ), given
f (t ), g (t ), w (t ) and x0 as
x (t ) = f (t )x (t ) + g (t )w (t ) with x (t0 ) = x0
The solution of this equation is
Rt Z t Rt
f ()d f ()d
x (t ) = e t0
x0 + e g ()w ()d
t0
Rt
f ()d
Suppose now we define (t , t0 ) = e t0
, we can write the
above solution as
Z t
x (t ) = (t , t0 )x0 + (t , )g ()w ()d
t0

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 33 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Mean of
a Linear First-Order Stochastic Process


Given a first-order stochastic process, (t ), with constant f
and g and white noise, w (t ), which is represented as

(
t ) = f (t ) + g w (t ) with (t0 ) = 0
and the mean and covariance of w (t ) expressed as

E [w (t )] = 0 and E [w (t )w ()] = q (t )
The mean of the process,
(t ) is
Rt Z t Rt
f d f d

(t ) = E [(t )] = e t0
0 + e g ()E [w ()]d
t0
Rt
f d
= e t0

0
= e f (t t0 )

0

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 34 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Stochastic Processes
The Mean-Square and

Covariance of a Linear First-Order Stochastic Process

The mean-square of the linear first-order stochastic process,

(t ) is
q h i
E [2 (t )] = e 2f (t t0 ) E [(t0 )(t0 )] + 1 e 2f (t t0 )
2f
q h i
= (t , t0 )E [(t0 )(t0 )] +
2
1 2 (t , t0 )
2f
The covariance of (t ), P (t ), is expressed as
h i
(t ))2
P (t ) = E ((t ) = E [2 (t )] 2 (t )
q h i
= 2 (t , t0 )P (t0 ) + 1 2 (t , t0 )
2f

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 35 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Stochastic Processes
The Vector First-Order

Differential Equation
A first-order vector differential equation for x(t ), given x(t0 ) and
white noise with E (w(t )) = 0, and E (w(t )w(T )) = Q (t ), is

x (t ) = F(t )x(t ) + G(t )w(t )

The solution of this equation is

Z t
x(t ) = (t , t0 )x(t0 ) + (t , )G()w()d
t0

where (t , t0 ) satisfies the following equation

t , t0 ) = F(t )(t , t0 ),
( with (t0 , t0 ) = I

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 36 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Mean and

Mean-Square of a Linear, Vector Process


The mean of the stochastic process x (t ) = F(t )x(t ) + G(t )w(t ) is
Z t
(t ) = E [x(t )] = (t , t0 )E [x(t0 )] +
x (t , )G()E [w()]d
t0
= (t , t0 )x
(t0 )
The mean-square of the process (with E [x(t0 )wT (t )] = 0) is
(" Z t #
T
E [x(t )x (t )] = E (t , t0 )x(t0 ) + (t , )G()w()d
t0
" Z t #)
(t , t0 )x(t0 ) + (t , )G()w()d
t0

= (t , t0 )E [x(t0 )x (t0 )]T (t , t0 )

T
Z t
+ (t , )G() Q GT ()T (t , ) d
t0

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 37 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Covariance
of a Linear, Vector Process


The covariance of x(t ), Pxx (t ), given Pxx (t0 ), is expressed as
h i
(t ))T
(t )) (x(t ) x
Pxx (t ) = E (x(t ) x = E [x(t )xT (t )] x T (t )
(t )x
= (t , t0 )Pxx (t0 )T (t , t0 )
Z t
+ (t , )G() Q GT ()T (t , ) d
t0

The differential equation for Pxx (t ) can be found to be

P xx (t ) = F(t )Pxx (t ) + Pxx (t )FT (t ) + G(t ) Q GT (t )

In the above development we have made use of the Sifting

Property of the Dirac Delta, (t ), expressed as
Z
f ()(t )d = f (t )

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 38 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

A Discrete Linear,
Vector Process

Given the continuous process (x (t ) = F(t )x(t ) + G(t )w(t )), whose
solution is
Z tk
x(tk ) = (tk , tk 1 )x(tk 1 ) + (t , )G()w()d
tk 1

the discrete stochastic analog process is

Z tk

xk = (tk , tk 1 )xk 1 + wk , with wk = (t , )G()w()d
tk 1

whose mean is

k = (tk , tk 1 )x
x k 1

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 39 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Covariance
of a Discrete Linear, Vector Process

Likewise, the continuous-time solution for the covariance was

Pxx (tk ) = (tk , t0 )Pxx (t0 )T (tk , t0 )

Z t
+ (tk , )G() Q GT ()T (tk , ) d
t0

whose discrete analog is

Pxxk = (tk , tk 1 )Pxxk 1 T (tk , tk 1 ) + Qk

where
Z t

Qk = (tk , )G() Q GT ()T (tk , ) d
t0

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 40 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Introduction and Background
Concepts from Probability Theory
Linear and Nonlinear Systems

Least Squares Estimation

The Kalman Filter

Stochastic Processes
The Kalman Filter Revealed

Implementation Considerations and Advanced Topics

The Extended Kalman Filter
Practical Considerations
Advanced Topics

Conclusions
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 41 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Practical Considerations

There is nothing more practical than a good theory

Albert Einstein

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 41 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Context of the Kalman Filter

With the advent of the digital computer and modern control,
the following question arose: Can we recursively estimate the
state of a vehicle as measurements become available?
In 1961 Rudolf Kalman came up with just such a methodology
to compute an optimal state given linear measurements and a
linear system
The resulting Kalman filter is an globally optimal linear,
model-based estimator driven by Gaussian, white noise which
has two steps
Propagation: the state and covariance are propagated from
one epoch to the next by integrating model-based dynamics
Update: the state and covariance are optimally updated with
measurements
We begin with the same equation as before

yk = Hk xk +  k with E ( k ) = 0, E ( k  Tk ) = Rk
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 42 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

What does a Kalman Filter do?

Fundamentally, a Kalman filter is nothing more than a

predictor (which we call the propagation phase) followed by a
corrector (which we call the update phase)
We use the dynamics (i.e. Newtons Laws) to predict the state
at the time of a measurement
The measurements are then used to correct or update the
predicted state.
It does this in an optimal fashion

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 43 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Prediction

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 44 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Measurement

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 45 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Compute Residual

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 46 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Correction

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 47 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Derivation
of the Kalman Filter (I)


k be an unbiased a priori estimate (the prediction) of xk with
Let x
covariance Pk so that the a priori estimate error, ek is
T
ek = xk x
k with E (ek ) = 0, E (ek ek ) = Pk
We hypothesize an unbiased linear update (the correction) to xk ,
+
called xk
, as follows (with Kk as yet unknown)
 
+
xk
=x k
k + Kk yk Hk x

whose a posteriori error, e+

k
, is
e+
k
+
= xk xk
= ek Kk (Hk ek +  k ) = (Ik Kk Hk )ek Kk 
If ek and  k are uncorrelated, then the a posteriori covariance is
T
P+
k
= E (e+ e+ ) = (I Kk Hk )Pk (I Kk Hk )T + Kk Rk K Tk
k k

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 48 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Derivation
of the Kalman Filter (II)


So far we havent said anything about Kk . We now choose Kk to
minimize the a posteriori error as1
1
 T  1   T  1   T 
min J = E e+k
e+k
= tr E e+ k
e+k
= E tr e+k
e+
k
2 2 2
1
  T
 1   T
 1  
= E tr e+ e+
k k
= tr E e+ e+
k k
= tr P+k
2 2 2
so we obtain K by2
  h i
tr P+ = tr (I Kk Hk ) P
( I K k Hk ) T
+ K k Rk K T
k = 0
Kk k Kk k

1
The cyclic invariance property of the trace is: tr(ABC) = tr(BCA) = tr(CAB)
2
Recalling that

tr(AXBXT ) = AT XBT + AXB; tr(AXB) = AT BT ; tr(AXT B) = BA
X X X
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 49 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Derivation
of the Kalman Filter (III)


This results in the following condition
 T  
Pk HTk Pk HTk + Kkopt Hk Pk HTk + Rk + Kkopt Hk Pk HTk + Rk = 0

which gives
  1
Kkopt = Pk HTk Hk Pk HTk + Rk

and substituting into the equation2 for P+ we get

  1  
P+
k
= Pk Pk HTk Hk Pk HTk + Rk Hk Pk = I Kkopt Hk Pk

so the state update is

 
+
xk
k + Kkopt yk Hk x
=x k

2
Recall that P+ = (I KH)P (I KH)T + KRKT
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 50 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Kalman Filter Revealed

Given the dynamics and the measurements
xk = (tk , tk 1 )xk 1 + k wk , with E (wk ) = 0, E (wk wTj ) = Qk kj
yk = Hk xk +  k , with E ( k ) = 0, E ( k  Tj ) = Rk kj
The Kalman Filter contains the following phases:
Propagation the Covariance Increases
k
x = (tk , tk 1 )x
+
k 1
Pk = (tk , tk 1 )P+
k 1
T (tk , tk 1 ) + k Qk Tk
Update the Covariance Decreases
  1
Kkopt = Pk HTk Hk Pk HTk + Rk
 
+
xk
k + Kkopt yk Hk x
= x k
     T
P+
k
= I Kkopt Hk Pk = I Kkopt Hk Pk I Kkopt Hk + Kkopt Rk KTkopt
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 51 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

A Kalman Filter
Example


Given a spring-mass-damper system governed by the following
equation
r (t ) = 0.001r (t ) 0.005r (t ) + w (t )
the system can be written (in first-order discrete form,
xk = (tk , tk 1 )xk 1 + k wk ) as
" # (" # )" # " #
r (tk ) 0 1 r (tk 1 ) 0
= exp t + wk
r (tk ) 0.001 0.005 r (tk 1 ) 1
with measurements
yk = r (tk ) + k with E [k ] = 0, E [j k ] = 0.0012 jk
and
" #
1 0
P0 = and Q = 0.0052
0 0.12
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 52 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

A Kalman Filter
Example (II)

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 53 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

A Kalman Filter
Example (III)

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 54 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Introduction and Background
Concepts from Probability Theory
Linear and Nonlinear Systems

Least Squares Estimation

The Kalman Filter

Stochastic Processes
The Kalman Filter Revealed

Implementation Considerations and Advanced Topics

The Extended Kalman Filter
Practical Considerations
Advanced Topics

Conclusions
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 55 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Practical Considerations

A computation is a temptation that should be resisted a long as

possible

John Boyd (paraphrasing T.S. Eliot) , 2000

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 55 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Extended
Kalman Filter

Since we live in a nonlinear and non-Gaussian world, can we fit the

Kalman filter paradigm into the real world? Being engineers,
when all else fails, we linearize.

Xk = X?
b
k + xk

This process results in an algorithm called the Extended Kalman

filter (EKF). However all guarantees of stability and optimality are
lost. The EKF is a conditional mean estimator with dynamics
truncated after first-order by deterministically linearizing about the
conditional mean.

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 56 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Development
of the Extended Kalman Filter (I)


Begin with the nonlinear state equation

X (t ) = f(X, t ) + w(t ) with E [w(t )] = 0, E [w(t )w()] = Q (t )

whose solution, given X(tk 1 ) is
Z t Z t
X(t ) = X(tk 1 ) + f(X, )d + w()d
tk 1 tk 1

We expand f(X, t ) in a Taylor series about X

= E (X) as

, t ) + f 
  
X(t ) = f(X X + + w(t )
X
X X=X

(t ), neglecting higher than first-order terms,
so that X
, t)
(t ) = f(X
X

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 57 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Development
of the Extended Kalman Filter (II)


 i
)(X X
)T h
Recalling the definition of P = E (X X , we find that

f 

P(t ) = F(t )P(t ) + P(t )FT (t ) + Q where F =
X X=X


which can be integrated as

P(tk ) = (tk , tk 1 )P(tk 1 )T (tk , tk 1 ) + Qk

with (tk 1 , tk 1 ) = I and

Z tk
t , tk 1 ) = F(t )(t , tk 1 ), and Qk =
( (tk , ) Q T (tk , )d
tk 1

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 58 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Development
of the Extended Kalman Filter (III)


Likewise, the measurement equation can be expanded in a Taylor
, the a priori state, as
series about X k
h 
  
Yk = h(Xk ) +  k = h(X
) + + + k
Xk X
k
Xk Xk =X k k


In the EKF development, we truncate the Taylor series after

first-order. As in the Kalman filter development, we minimize the
trace of the a posteriori covariance and this results in
h i 1
) = P HT (X
Kk (X ) Hk (X
) P HT (X
) + Rk
k k k k k k k k
h i
P+ = )HT (X
I Kk (X ) P
k k k k k
h i
+
X + Kk (X
= X ) Yk hk (X
)
k k k k


) = h
Hk (X

k
Xk 

Xk =Xk
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 59 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

The Extended
Kalman Filter (EKF) Algorithm

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 60 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Introduction and Background
Concepts from Probability Theory
Linear and Nonlinear Systems

Least Squares Estimation

The Kalman Filter

Stochastic Processes
The Kalman Filter Revealed

Implementation Considerations and Advanced Topics

The Extended Kalman Filter
Practical Considerations
Advanced Topics

Conclusions
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 61 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Practical Considerations

In theory, there is no difference between theory and practice, but

in practice there is

John Junkins, 2012

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 61 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Practical Matters
Processing Multiple Measurements


In general, more than one measurement will arrive at the
same time
Usually the measurements are uncorrelated and hence they
can be processed one-at-a-time
However, even if they are correlated, they can usually be
treated as if they were uncorrelated by increasing the
measurement noise variance
If the measurements are processed one-at-a-time, then
 1 Pk HTk
Kk = Pk HTk Hk Pk HTk + Rk =
Hk Pk HTk + Rk
Thus there is no need for a matrix inverse we can use scalar
division
This greatly reduces the computational throughput, not

to
mention software complexity
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 62 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Practical Matters
Processing Non-Gaussian

Measurements

The Kalman Filter is predicated on measurements whose

errors are Gaussian
However, real-world sensors seldom have error
characteristics that are Gaussian
Real sensors have (significant) biases
Real sensors have significant skewness (third moment) and
kurtosis (fourth moment)
A great deal of information is contained in the tails of the
distribution
Significant sensor testing needs to be performed to fully
characterize a sensor and determine its error characteristics
Measurement editing is performed on the innovations process
(ik = Yik hi (X
) with variance Vi = Hi P HT + Ri )
k k k k ik k
Dont edit out measurements that are greater than 3V ik
We process measurements that are up to 6Vik
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 63 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Practical Matters
Dealing with Measurement Latency


Measurements arent so polite as to be time-tagged or to
arrive at the major cycle of the navigation filter (tk )
Therefore, we need to process the measurements at the time
they are taken, assuming that the measurements are not too
latent
Provided they are less than (say) 3 seconds latent
The state is propagated back to the measurement time using,
say, a first-order integrator
f
Xm = Xk + f(Xk )t + (Xk )f(Xk )t 2
X
The measurement partial mapping is done in much the same
way as it was done in batch estimation
Map the measurement sensitivity matrix at the time of the
measurement(H(Xm )) to the filter time (tk ) using the state

transition matrix, (tm , tk ).
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 64 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Practical Matters
Measurement Underweighting


Sometimes, when accurate measurements are introduced to
a state which isnt all that accurate, filter instability results
There are several ways to handle this
Second-order Kalman Filters
Sigma Point Kalman Filters
Measurement Underweighting
Since Apollo, measurement underweighting has been used
extensively
What underweighting does is it slows down the rate that the
measurements decrease the covariance
It approximates the second-order correction to the covariance
matrix
Underweighting is typically implemented as
  1
Kk = Pk HTk (1 + )Hk Pk HTk + Rk
The scalar is a tuning parameter used to get good filter
performance
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 65 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Practical Matters
Filter Tuning (I)

Regardless of how you slice it, tuning a navigation filter is an

art
There are (at least) two sets of knobs one can turn to tune a
filter
Process Noise (also called State Noise or Plant Noise), Q,
the noise on the state dynamics
Measurement Noise, R
Filter tuning is performed in the context of Monte Carlo
simulations (1000s of runs)
FIlter designers begin with the expected noise parameters
Process Noise the size of the neglected dynamics (e.g. a
truncated gravity field)
Measurement Noise the sensor manufacturers noise

specifications

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 66 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Practical Matters
Filter Tuning (II)


Sensor parameters (such as bias) are modeled as zero-mean
Gauss-Markov parameters, xp , which have two tuning
parameters
The Steady State Variance (Ppss )
The Time Constant ()
d 1
xp = xp + wp , where E [wp (t )wp ()] = Qp (t )
dt p
Ppss
Qp = 2
p
All of these are tuned in the Monte Carlo environment so that
The state error remains mostly within the 3- bounds of the
filter covariance
The filter covariance represents the computed sample

covariance
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 67 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Practical Matters
Filter Tuning (III)

Sometimes the filter designer inadvertently chooses a process

noise such that the covariance of the state gets too small
When this happens, the filter thinks it is very sure of itself it
is smug
The end result is that the filter starts rejecting measurements
Never a good thing
The solution to this problem is to inject enough process noise
to keep the filter open
This allows the filter to process measurements appropriately
There are several spacecraft which have experienced
problems because the designers have chosen incorrect (too
small) process noise
Of course, this is nothing more than the classic tension

between stability and performance
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 68 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Practical Matters
Invariance to Measurement Ordering

Because of its nonlinear foundation, the performance of an

EKF can be highly dependent on the order in which
measurements are processed
For example, if a system processes range and bearing
measurements, the performance of the EKF will be different if
the range is processed first versus if the bearing were
processed first
To remedy this, on Orion we employ a hybrid linear/EKF
formulation
The state and covariance updates are accumulated in delta
state and covariance variables
The state and covariance are updated only after all the
measurements are processed

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 69 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Introduction and Background
Concepts from Probability Theory
Linear and Nonlinear Systems

Least Squares Estimation

The Kalman Filter

Stochastic Processes
The Kalman Filter Revealed

Implementation Considerations and Advanced Topics

The Extended Kalman Filter
Practical Considerations
Advanced Topics

Conclusions
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 70 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Advanced Topics

The Kalman-Bucy Filter

The Schmidt-Kalman Consider Filter
The Kalman Smoother
Square Root and Matrix Factorization Techniques
Potter Square Root Filter (Apollo)
UDU Filter (Orion)
Nonlinear Filters
Second-Order Kalman Filters
Sigma Point Kalman Filters
Particle Filters
Entropy Based / Bayesian Inference Filters
Linear Covariance Analysis

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 70 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Introduction and Background
Concepts from Probability Theory
Linear and Nonlinear Systems

Least Squares Estimation

The Kalman Filter

Stochastic Processes
The Kalman Filter Revealed

Implementation Considerations and Advanced Topics

The Extended Kalman Filter
Practical Considerations
Advanced Topics

Conclusions
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 71 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

Conclusions



Kalman Filtering and Least Squares Estimation are at the
heart of the spacecraft navigation
Ground-based navigation
On-board navigation
Its purpose is to obtain the best state of the vehicle given a
set of measurements and subject to the computational
constraints of flight software
It requires fluency with several disciplines within engineering
and mathematics
Statistics
Numerical Algorithms and Analysis
Linear and Nonlinear Analysis
Sensor Hardware
Challenges abound
Increase demands on throughput
Image-based sensors
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 71 / 73
 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

To put things
in perspective

I never, never want to be a pioneer . . . Its always best to come in

second, when you can look at all the mistakes the pioneers made
and then take advantage of them.
Seymour Cray

Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 72 / 73

 Introduction and Background
Least Squares Estimation
The Kalman Filter NASA / Johnson Space Center
Implementation Considerations and Advanced Topics

References

Maybeck, P.S. Stochastic Models, Estimation, and Control,

Volumes 1-3, Academic Press, 1982.
Gelb, A. E. (ed.), Applied Optimal Estimation, MIT Press,
1974.
Brown, R.G., and Hwang, P.Y.C.,Introduction to Random
Signals and Applied Kalman Filtering, John Wiley, 1992.
Simon, D., Optimal State Estimation: Kalman, H-infinity, and
Nonlinear Approaches, John WIley, 2006.
Tapley, B., Schutz, B., and Born, G., Statistical Orbit
Determination, Elsevier Inc, 2004.
Zanetti, R., DeMars, K, and Bishop, R., Underweighting
Nonlinear Measurements, Journal of Guidance, Control, and

Dynamics, Vol. 33, No. 5, September-October 2010.
Christopher DSouza March 2013 Fundamentals of Kalman Filtering and Estimation 73 / 73