Kalman Filtering

It is an effective and versatile mathematical procedure for

combining noisy sensor outputs to estimate the state of a
system with uncertain dynamics.
Kalman filtering is a relatively recent (1960) development in
filtering.
Kalman filtering has been applied in areas as diverse as
aerospace, tracking missiles, navigation, nuclear power
plant instrumentation, demographic modeling,
manufacturing, computer vision applications.
For Kalman filter the problem is formulated is state space
and is time varying.

What is Kalman filter

Less noisy data

Noisy data
Some Applied Math

Delay is the price for filtering

Introduction

The Kalman filter is a linear, recursive estimator that

produces the minimum variance estimate in a least squares
sense under the assumption of white, Gaussian noise processes.
Consider the problem of estimating the variables of a system. In
dynamic systems (that is, systems which vary with time) the system
variables are often denoted by the term "state variables".
Since its introduction in 1960, the Kalman filter has become an
integral component in thousands of military and civilian navigation
systems.

This deceptively simple, recursive digital algorithm has been an

early-on favorite for conveniently integrating (or fusing) navigation
sensor data to achieve optimal overall system performance.
3

The Kalman filter is a multiple-input, multiple-output digital filter

that can optimally estimate, in real time, the states of a system
based on its noisy outputs.

The purpose of a Kalman filter is to estimate the state of a system

from measurements which contain random errors. An example is
estimating the position and velocity of a satellite from radar data.
There are 3 components of position and 3 of velocity so there are at
least 6 variables to estimate. These variables are called state
variables. With 6 state variables the resulting Kalman filter is called
a 6-dimensional Kalman filter.

To provide current estimates of the system variables - such as

position coordinates - the filter uses statistical models to properly
weight each new measurement relative to past information.
4

Predict Correct

Kalman filter operates by

Predicting the new state and its uncertainty
Correcting with the new measurement.

Predict

Correct

Example: Two-dimensional Position Only

Process Model

xk 1 0 xk 1 ~ xk 1
y 0 1 y ~ y
k 1
k 1
k
State

State transition

Nose

xk Axk 1 wk 1 Or
xk 1 Ak xk wk
Noise

Measurement Model

uk H x 0

0
H
v
y

measurement

xk

~ uk

y k ~ vk

Measurement matrix

Nose

z k Hxk vk

Noise

Preparation

1
A
0

1 State Transition

0
Q xx
T
Q E w w
Process Noise Covariance
Q yy
0

R xx
T
R E v v
0

Measurement Noise Covariance

R yy

Initialization

xo Hzo

Po
0

Predict

xk Axk 1

T
Pk APk 1 A Q

Uncertainty

Transition

10

Correct

xk xk K z k Hxk

Actual

Predicted

Pk I KH Pk

K Pk H HPk H R

Measurement space

11

Summary

xk Axk 1
Pk APk 1 AT Q

K Pk H T HPk H T R

xk xk K ( z k Hxk )
Pk I KH Pk

12

Review of State Model

xk 1 Ak xk wk
z k H k xk vk

White noise

xk (n n) state vector at time t k

Ak (n n) state transition matrix
wk (n 1) forcing function
z k (m 1) measurement vector
H k (m m) output matrix
vk (m 1) measurement noise
13

The System
Measurement noise
vk
+

wk

xk+1

z-1

xk

Hk

zk

+
Ak
State transition matrix

14

Kalman Predictor Update Equation

xk1 Axk A xk K z k Hxk

Prior estimate of xk

The predictor will turn out to be a system that looks much like
the actual system.

15

The Predictor

+
zk

Ak K

xk1

xk

Z-1

zk

Ak

Ak

16

The Filter

Zk+1

Ak K

xk1

xk

Z-1

zk

Ak

Ak

17

How Kalman Filter Works?

The Kalman filter maintains two types of variables:

Estimate State Vector: The components of the estimated state
vector include the following:
The variables of interest (what we want or need to know, such as
position and velocity).
Nuisance variables that may be necessary to the estimation process.
The Kalman filter state variables for a specific application must include
all those system dynamic variables that are measurable by the sensors
used in the application.

A Covariance Matrix: a measure of estimation uncertainty. The

equations used to propagate the covariance matrix (collectively
called the Riccati equation) model and manage uncertainty, taking
into account how sensor noise and dynamic uncertainty contribute
to uncertainty about the estimated system state.
18

By maintaining an estimate of its own estimation uncertainty and the

relative uncertainty in the various sensor outputs, the Kalman filter is
able to combine all sensor information optimally in the sense that
the resulting estimate minimizes any quadratic loss function of
estimation, including the mean-squared value of any linear
combination of static estimation errors.

The Kalman gain is the optimal weighting matrix for combining new
sensor data with a prior estimate to obtain a new estimate.

19

Lecture 1: The Start

For the Kalman filter, the problem is formulated in state space.

Consider a linear system: u(t) and y(t) could be scalars or vectors.
Each element of u(t) is a white noise.
We want to model y(t) as the response of a linear system, where the system
input is the unity power spectrum white noise u(t). This implies E [x(t)] = 0.

x Fx Gu
y Bx
u(t)

Linear System

y(t)

20

Suppose we have a fourth order system

Assume n = 4, then m = n -1 = 3
Based on the information given in Chapter 3 of the State Space we can
write the matrix equation

Forward path

u(t)

bm s m bm 1s m 1 .. bo
s n an 1s n 1 .. ao

y(t)

Feedback path

21

b3

b1

b2
1/s

1
U(s)

G(s)

-a3

1/s
x4

1/s
x3

-a2

b3 s 3 b2 s 2 b1s b0
s 4 a3 s 3 a 2 s 2 a1s a0
b3 s 1 b2 s 2 b1s 3 b0 s 4

1 a3 s 1 a 2 s 2 a1s 3 a0 s 4

Read Examplev 3.1 of the textbook

-a1

1/s
x2

-a0

bo
x1

Y(s)

.
.
.
x1 x2 ; x 2 x3 ; x 3 x4
.
x 4 a0 x1 a1x2 a2 x3 a3 x4 u
y (t ) b0 x1 b1x2 b2 x3 b3 x4
x1 0

d x2 0

0
dt x3

x4 - a0

1
0
0
- a1

0
0
1
0
0
1
- a2 - a3

x1 0
x
2 0 u (t )
x3 0


x4 1

x1

x
2

y (t ) b0 b1 b2 b3
x3

x4

22

Discrete-Time State Space Model

In the Kalman filtering it is customary to write w(k) and y(k)
We will repeat the process as before

Forward path

u(t)

m s m m 1s m 1 ..
s n n 1s n 1 .. o

y(t)

Feedback path

23

b3

b1

b2
1/s

1
U(s)

-a3

1/s
x4

1/s
x3

-a2

3s 3 2 s 2 1s 0
G( s)
s 4 3s 3 2 s 2 1s 0
3 s 1 2 s 2 1s 3 0 s 4

1 3 s 1 2 s 2 1s 3 0 s 4

Read Examplev 3.1 of the textbook

-a1

1/s
x2

-a0

bo
x1

Y(s)

.
.
.
x1 x2 ; x 2 x3 ; x 3 x4
.
x 4 0 x1 1x2 2 x3 3 x4 u
y (t ) 0 x1 1x2 2 x3 3 x4
1 0
x1 (k 1) 0

0
1
d x 2 ( k 2) 0

0
0
dt x3 (k 3) 0

x4 (k 1) - 0 - 1 - 2
x1 (k )

x
(
k
)
2

y (t ) 0 1 2 3
x3 (k )

x4 (k )

x1 (k ) 0
x ( k )
2 0 w(t )
x3 (k ) 0
1


- 3 x4 (k ) 1
0
0

24

Writing the former equations together using matrix notation we obtain

the controllable state variable

X k 1 k X k Wk

Yk B k X k
X k State vector at time k
k State transition matrix for X k X k 1
B k Output equation gain matrix
E[X k ] 0

25

Development of the Discrete Kalman Filter

There should be a discrete linear system.

The input is white noise.
The observations are the system output plus a white noise called
the measurement noise.
The system input noise and the measurement noise are
uncorrelated to each other.
You should know:
The state space model for the system.
The second order statistics of the input noise.
The second order statistic of the measurement noise.

The problem: Given the noisy observations of the output, find

estimates of the system state vector.

26

What Makes Kalman Filter Different?

It is kind of like a mathematical proof by induction.

Assume that we have obtained a prediction for the state vector at time k and
that this estimate is based on the first k-1 observations.
In other words, assume that we have an estimate of Xk given Zk-3, Zk-2, Zk-1.
This is called a priori estimate or prior of Xk: the true state vector at time k.
In books or other resources you may see it as

X
k
- E X so the error in the prediction X
- is
E X
k
k
k
- (X true; X
- predicted)
e -k X k X
k
k
k
-
with mean E e -k E X k - X
k

e k is the predicted state vector error.

27

The Predicted State Vector Error Covariance Matrix

Pk- E e -k e T E
k

X - X
-
k
k

X X
-
k
k

X k I - K k H k X k K k Zk
X
- K (Z H X
- )
X
k
k
k k
k k

28

Lecture 2: State and Covariance Correction

The Kalman filter is a two-step process: prediction and
correction.
The filter can start with either step but will begin by describing the
correction step first.
The correction step makes corrections to an estimate, based on
new information obtained from sensor measurements.
The Kalman gain matrix K is the crown jewel of Kalman filter. All
the efforts of solving the matrix is for the sole purpose of
computing the optimal value of the gain materix K used for
correction of an estimate x .

x ( ) x () K gain z Hx ( )
Based on measurement

z Hx noise
29

There are two basic processes that are modeled by a Kalman filter. The first process
is a model describing how the error state vector changes in time. This model is the
system dynamics model. The second model defines the relationship between the
error state vector and any measurements processed by the filter and is the
measurement model.

Time Update

Measurement Update
(Correct)

30

Intuitively, the Kalman filter sorts out information and

weights the relative contributions of the measurements and of
the dynamic behavior of the state vector.

The measurements and state vector are weighted by their

respective covariance matrices. If the measurements are inaccurate
(large variances) when compared to the state vector estimate, then
the filterwill deweight the measurements.

On the other hand, if the measurements are very accurate (small

variances)
when
compared
to the state estimate, then the filter will tend to weight the
measurements heavily with the consequence that its previously
computed state estimate will contribute little to the latest
state estimate.
31

Gaussian Probability Density Function

PDFs are nonnegative integrable functions whose integral equals
unity. The density function of Gaussian probability distributions have
the form given. Where n is the dimension of P (nn matrix), is the
mean of the distribution. The parameter P is the covariance matrix
of the distribution.

1
T
p( x)
exp( x P 1[ x ])
2
(2 ) n det P

32

Likelihood Functions

Likelihood functions are similar to probability density functions, except

that their integrals are not constrained to equal unity, or even required to
be infinite. They are useful for comparing relative likelihoods and for
finding the value of the unknown independent variable x at which the
likelihood function achieves its maximum.
Y is called the information matrix of the likelihood function. It replaces P-1
in the Gaussian probability density function. If the information matrix Y is
nonsingular, then its inverse Y-1 = P.

1
T
(x, , Y) exp( [x ] Y[x ])
2

33

The purpose of a Kalman filter is to optimally estimate the values of variables describing
the state of a system from a multidimensional signal contaminated by noise

Multiple
noise
inputs

System: Unknown multiple state variables

Sampled multiple output

Multiple noises

Multiply noisy outputs

Multidimensional signal plus noise

Kalman filter

Multiple state
Variable estimates
34

 The following figure illustrates the Kalman filter algorithm itself.

Because the state (or signal) is typically a vector of scalar random
variables (rather than a single variable), the state uncertainty
estimate is a variance-covariance matrix-or simply, covariance
matrix. Each diagonal term of the matrix is the variance of a scalar
random variable-a description of its uncertainty. The term is the
variable's mean squared deviation from its mean, and its square
root is its standard deviation. The matrix's off-diagonal terms are the
covariances that describe any correlation between pairs of
variables.

The multiple measurements (at each time point) are also vectors
that a recursive algorithm processes sequentially in time. This
means that the algorithm iteratively repeats itself for each new
measurement vector, using only values stored from the previous
cycle. This procedure distinguishes itself from batch-processing
algorithms, which must save all past measurements.
35

 Starting with an initial predicted state estimate (as shown in the

figure) and its associated covariance obtained from past
information, the filter calculates the weights to be used when
combining this estimate with the first measurement vector to obtain
an updated "best" estimate. If the measurement noise covariance is
much smaller than that of the predicted state estimate, the
measurement's weight will be high and the predicted state
estimate's will be low.
Also, the relative weighting between the scalar states will be a
function of how "observable" they are in the measurement. Readily
visible states in the measurement will receive the higher weights.
Because the filter calculates an updated state estimate using the
new measurement, the state estimate covariance must also be
changed to reflect the information just added, resulting in a reduced
uncertainty. The updated state estimates and their associated
covariances form the Kalman filter outputs.

36

Finally, to prepare for the next measurement vector, the filter must
project the updated state estimate and its associated covariance to
the next measurement time.

The actual system state vector is assumed to change with time

according to a deterministic linear transformation plus an
independent random noise.

Therefore, the predicted state estimate follows only the deterministic

transformation, because the actual noise value is unknown. The
covariance prediction ac-counts for both, because the random
noise's uncertainty is known.

Therefore, the prediction uncertainty will increase, as the state

estimate prediction cannot account for the added random noise.
This last step completes the Kalman filter's cycle.

37

Compute weights from predicted states

Covariance and measurement noise covariance

Predicted
initial state
Estimate and covariance

Predict state estimates and

Covariance to next time step

New measurements each

cycle
Update state estimates as
Weighted linear blend of predicted
state estimates & new measurement

Updated state estimates

Compute new covariance of updated state estimates

38

Mathematical Definitions
The variance and the closely-related standard deviation
are measures of how spread out a distribution is. It is a
measure of estimation quality.
The covariance is a statistical measure of correlation of
the fluctuations of two different quantities. Intuitively,
covariance is the measure of how much two variables
vary together.
Least squares is a mathematical optimization technique
which, when given a series of measured data, attempts
to find a function which closely approximates the data (a
"best fit"). It attempts to minimize the sum of the squares
of the ordinate differences (called residuals) between
points generated by the function and corresponding
points in the data. It is sometimes called least mean
squares.
39

Simple Example of Process Model

A simple hypothetical example may help clarify the
Kalman concepts. Consider the problem of determining
the actual resistance of a nominal 100-ohm resistor by
making repeated ohmmeter measurements and
processing them in a Kalman filter.
First, one must determine the appropriate statistical
models of the state and measurement processes so that
the filter can compute the proper Kalman weights (or
gains). Here, only one state variable, the resistance, x is
unknown but assumed to be constant. So the state
process dynamics evolves with time as

Xk+1 = Xk.

[1]
40

Note that no random noise corrupts the state process as it evolves

with time. The color code on a resistor indicates its precision, or
tolerance, from which one can deduce assuming that the population
of resistors has a Gaussian or normal histogram that the uncertainty
(variance) of the 100-ohm value is, say, (2 ohm)2. So our best
estimate of x, with no measurements, is x0/ = 100 with an
uncertainty of P0/ = 4. Repeated ohmmeter measurement,

zk = xk + vk

[2]

directly yield the resistance value with some measurement noise, vk

(measurement errors from turn-on to turn-on are assumed
uncorrelated). The ohmmeter manufacturer indicates the
measurement noise uncertainty to be Rk = (1 ohm)2 with an average
value of zero about the true resistance.

41

How it Works?

Estimated State Vector including the variables of interest, nuisance

variables, and the Kalman filter state variables for a specific
application.

A Covariance Matrix: A measure of estimation uncertainty. The

equations used to propagate the covariance matrix (called Riccatti
equation) model and manage uncertainty, taking into account how
sensor noise and dynamic uncertainty contribute to uncertainty
about the estimated system state.

Kalman filter is able to combine all sensor information optimally in

the sense that the resulting estimate minimizes any quadratic loss
function of estimation error.

The Kalman gain is the optimal weighting matrix for combining new
sensor data with a prior estimate to obtain a new estimate.
42

For our purpose in ELG4152

The noisy sensors may include speed sensors (wheel speeds of

land vehicles, water speed sensors for ships, air speed sensors for
aircraft, GPS receivers and inertia sensors, and time sensors.

The system state may include the position, velocity, acceleration,

attitude, and attitude rate of a vehicle on land, at sea, in the air, or in
the space.

Uncertain dynamics may include unpredictable disturbances of the

host vehicle, whether caused by a human operator or by the
medium (winds, surface currents, turns in the road, or terrain
changes). It might include also unpredictable changes in the sensor
parameters.

43

The one dimensional Kalman Filter

Suppose we have a random variable x(t) whose value we want to
estimate at certain times t0 ,t1, t2, t3, etc. Also, suppose we know that
x(tk) satisfies a linear dynamic equation.

x(tk+1) = Fx(tk) + u(k) (the dynamic equation)

In the above equation F is state transition matrix (in this example a

known number) that relates state at time step tk to time step tk+1. In
order to work through a numerical example let us assume F = 0.9.
Kalman assumed that u(k) is a random number. Suppose the
numbers are such that the mean of u(k) = 0 and the variance of
u(k) is Q. For our numerical example, we will take Q to be 100.

u(k) is called white noise, which means it is not correlated with any
other random variables and most especially not correlated with past
values of u.
44

In later lessons we will extend the Kalman filter to cases where the
dynamic equation is not linear and where u is not white noise. But
for this lesson, the dynamic equation is linear and w is white noise
with zero mean.

Now suppose that at time t0 someone came along and told you he
thought x(t0) = 1000 but that he might be in error and he thinks the
variance of his error is equal to P. Suppose that you had a great
deal of confidence in this person and were, therefore, convinced
that this was the best possible estimate of x(t0). This is the initial
estimate of x. It is sometimes called the a priori estimate.

A Kalman filter needs an initial estimate to get started. It is like an

automobile engine that needs a starter motor to get going. Once it
gets going it does not need the starter motor anymore. Same with
the Kalman filter. It needs an initial estimate to get going. Then it will
not need any more estimates from outside.
45

We have an estimate of x(t0),which we will call xe. For our example

xe = 1000. The variance of the error in this estimate is defined by P
= E [(x(t0) -xe)2].

where E is the expected value operator. x(t0) is the actual value of x

at time t0 and xe is our best estimate of x. Thus the term in the
parentheses is the error in our estimate. For the numerical example,
we will take P = 40,000.

Now we would like to estimate x(t1). Remember that the first

equation we wrote (the dynamic equation) was

x(tk+1) = Fx(tk) + u(k).

Therefore, for k = 0 we have x(t1) = Fx(t0) + u(0).

Dr. Kalman says our new best estimate of x(t1) is given by

New xe = Fxe (Eq. 1) or in our numerical example 900.

46

 We have no way of estimating u(0) except to use its mean value of

zero. How about Fx(t0). If our initial estimate of x(t0) = 1000 was
correct then Fx(t0) would be 900. If our initial estimate was high,
then our new estimate will be high but we have no way of knowing
whether our initial estimate was high or low (if we had some way of
knowing that it was high than we would have reduced it). So 900 is
the best estimate we can make. What is the variance of the error of
this estimate?

New P = E [(x(t1) new xe)2]

Substitute the above equations in for x(t1) and new xe and you get

New P = E [(Fx(t0) + u - Fxe)2]

= E [F2(x(t0) - xe)2 ] + E u2 + 2F E (x(t0)- xe)*u]

The last term is zero because u is assumed to be uncorrelated with

x(t0) and xe.
47

So, we are left with

New P = PF2 + Q (Eq. 2)
For our example, we have
New P = 40,000 X .81 + 100 = 32,500
Now, let us assume we make a noisy measurement of x. Call the
measurement y and assume y is related to x by a linear equation.
(Kalman assumed that all the equations of the system are linear.
This is called linear system theory.
y(1) = Mx(t1) + w(1)
where w is white noise. We will call the variance of w, "R".
M is some number whose value we know. We will use for our
numerical example M = 1 , R = 10,000 and y(1) = 1200
Notice that if we wanted to estimate y(1) before we look at the
measured value we would use

48

ye = M*new xe
for our numerical example we would have ye = 900
Dr. Kalman says the new best estimate of x(t1) is given by
Newer xe = new xe + K*(y(1) - M*new xe)
= new xe + K*(y(1) - ye) (Eq. 3)
where K is a number called the Kalman gain.

Notice that y(1) - ye is just our error in estimating y(1). For our
example, this error is equal to plus 300. Part of this is due to the
noise, w and part to our error in estimating x.
If all the error were due to our error in estimating x, then we would
be convinced that new xe was low by 300. Setting K = 1 would
correct our estimate by the full 300. But since some of this error is
due to w, we will make a correction of less than 300 to come up with
newer xe. We will set K to some number less than one.

49

 What value of K should we use? Before we decide, let us compute

the variance of the resulting error

E (x(t1) newer xe)2 = E [x new xe - K(y - M new xe)]2

= E [(x new xe - K(Mx + w - M new xe)]2

= E [{(1 - KM) (x new xe)2 +Kw}]2

= new P(1 - KM)2 + RK2

The cross product terms dropped out because w is assumed

uncorrelated with x and new xe. The newer value of the variance is
Newer P = new P (1 - KM)2 + R(K2) (Eq. 5)
If we want to minimize the estimation error we should minimize
newer P. We do that by differentiating newer P with respect to K and
setting the derivative equal to zero and then solving for K. A little
algebra shows that the optimal K is given by

K = M new P/ [new P(M2) + R] (Eq. 4)

For our example, K = .7647 ; Newer xe = 1129; newer P = 7647

Notice that the variance of our estimation error is decreasing.
50

These are the five equations of the Kalman filter. At time t2, we start
again using newer xe to be the value of xe to insert in equation 1
and newer P as the value of P in equation 2.

Then we calculate K from equation 4 and use that along with the
new measurement, y(2), in equation 3 to get another estimate of x
and we use equation 5 to get the corresponding P. And this goes on
computer cycle after computer cycle.
In the multi-dimensional Kalman filter, x is a column matrix with
many components. For example if we were determining the orbit of
a satellite, x would have 3 components corresponding to the
position of the satellite and 3 more corresponding to the velocity
plus other components corresponding to other random variables.
Equations 1 through 5 would become matrix equations and the
simplicity and intuitive logic of the Kalman filter becomes obscured.

51

Kalman Filter Applications: GPS/Automobile

Display

i
DGPS
Reference
Station

DGPS
Receiver

Radio
Modem
Trans

Radio
Modem
Receiver

Kalman Filter

Wheel
Speed
Sensor

Magnetic
Compass

Map Match System state vector

could include 10 state
Variables: 3 position;
Three velocity; receiver
Clock bias;
wheel speed;
Magnetic Compass.
Roadmap
Database

Can be added to the system state vector

52

GPS/Land Vehicle Integration

State Vector: 10 state variables

Measurement Vector: Inputs to the Kalman filter include the following:
Pseudoranges i to each satellite being tracked by the receiver.
Integrated Doppler components i for each of the satellites.
From the wheel speed sensor: indicated vehicle speed.
From the magnetic compass: magnetic heading angle.

vE
v h , , S wheel , S wheelspeed v
N
(1 S wheelspeed ) S wheel

sin( - )

v
(1

S
)
S
cos(

wheelspeed
wheel
vE east velocity; v N north velocity; v sensor noise

magnetic compass output angle; magnetic

compass output bias; S wheel wheel speed sensor output;
Swheelspeed wheel speed sensor scale factor offset.
The measurement sensitivity matrix for these variables is the partial
derivative of h evaluated at the estimated values of the state variables.

53

GPS / INS Integration

Kalman filter provides a simple algorithm that can easily lend itself to
integrated systems and requires only adequate statistical models of
the state variables and associated noises for its optimal
performance.

Integrating GPS with an inertial navigation system (INS) and a

Kalman filter provides improved overall navigation performance.
Essentially, the INS supplies virtually noiseless outputs that slowly
drift off with time. GPS has minimal drift but much more noise. The
Kalman filter, using statistical models of both systems, can take
advantage of their different error characteristic to optimally minimize
their deterious traits.

54

Integrated GPS Receiver and Inertial Navigator

GPS
Receiver

Measurement error

True + GPS error

Kalman
Filter
Inertial error
estimates
Predict
GPS

Inertial
Navigator

True + Inertial errors

Best total
estimates

55

Case Study

Write an article (case study) describing a consideration or

application based on Kalman filter. You may make use of the control
tool box of MATLAB.

56