You are on page 1of 11

Ulrik Sderstrm

ulrik.soderstrom@tfe.umu.se
9 Dec 2005

Kalman filter

General Introduction

What is a Kalman filter


z
z

Optimal recursive data processing algorithm


Typical Kalman filter application

Concept introduction
z

Kalman filter
z
z
z

A set of mathematical equations


Iterative, recursive process
Optimal data processing algorithm under certain
criteria
z

Discrete linear data

Extended KF, Unscented KF

Estimates past, present, future states

Optimal
z

Dependent upon the criteria chosen to


evaluate performance
Under certain assumptions, KF is optimal
with respect to virtually any criteria that
makes sense
z
z

Linear data
Gaussian model

Mean = Median = Max

Recursive
z

A Kalman filter only needs info from the


previous state
z
z

Updated for each iteration


Older data can be discarded
z

Saves computation capacity and storage

Variable definition
z
z

xk = state vector, process to examine


wk = process noise
z

vk = measurement noise
z
z

z
z
z
z

White, Gaussian, mean=0, Covariance matrix Q


White, Gaussian, mean=0, Covariance matrix R
Uncorrelated with wk

Sk = Covariance of the innovation, residual


Kk = Kalman gain, gain matrix
Pk = Covariance of prediction error
zk = Measurement of system state
7

Mathematics
z

Xk+1=Axk + wk

posk +1 1 0 posk t 2 / 2
ak
vel = 0 1 vel +
t
k
k +1

zK = posk + vk

More Mathematics
S k = Pk + R
K k = APk S k1
Pk +1 = APk AT + Q APk S k1 Pk AT
xk +1 = Axk + K k ( z k +1 Axk )

Kalman gain
z
z

Kk in the previous slides


Relates the new estimate to the most certain
of the previous estimates
z
z

Large measurement noise -> Small gain


Large system noise -> Large gain

System and measurement noise unchanged


z

Steady-state Kalman Filter

10

Iterative calculations
z

Prediction
z
z

The state
The error covariance

Initial
estimates
Update

Update
z
z
z

Kalman gain
Update with new measurement
Update with new error covariance

Predict

11

Iterative calculations
z

Prediction
z
z

xk=Axk-1 + wk-1

Pk = APk 1 AT + Q APk 1S k11 Pk 1 AT

Update
1
z K = AP S
k
k k
k +1 = Axk + K k ( z k +1 Axk )
z x
z P
k +1 = ( I K k S k ) Pk

Initial
estimates

Update

Predict
12

Conclusions
z

Given
z

A system model
z
z

With random initial state


Driven by white noise

Noisy measurements of linear combinations of the


system state variables

Determine
z

The best estimate of the system state variables

13

Applications

Anything that can be measured and


estimated
z
z

Initial state, noise models, system model


Often difficult to estimate system model
z

z
z
z

Which type of KF should be used?

Estimation
Filtering
Prediction
15

Tracking in VR
z

Pose
z

Where is somebody looking

Head position (any limb)


z

Where is the limb?

16

Navigation
z

Space shuttles
z

Keep shuttle stable

GPS
z

Received GPS signal can be refined with a


Kalman filter

17

Control system
z

Nuclear plants
z
z

Large system
Important to know the system state

Climate control systems


z

Cars, houses

18

Image processing
z

Image compression
z
z

Update compression iteratively


Resembling Matching Pursuit

Image filtering
z
z

Un-linear filtering
Remove certain features

19

Economics
z

Large scale models


z
z
z

Little control
Many parameters
Not correct models

20

10

Fridays presentation
z

MSP 1
z
z

Discrete Kalman filter


Extended Kalman filter

MSP 2
z
z

Unscented Kalman filter


Comparison of EKF and UKF

21

11

You might also like