Vous êtes sur la page 1sur 10

# Robot Mapping KF, EKF and UKF

## ! Kalman filter requires linear models

! EKF linearizes via Taylor expansion
Unscented Kalman Filter
Is there a better way to linearize?

1 2

## Linearization of the non-linear Compute a set of (so-called)

function through Taylor expansion sigma points

3 4
Unscented Transform Unscented Transform

## Transform each sigma point Compute Gaussian from the

through the non-linear function transformed and weighted points

5 6

## Unscented Transform Overview Sigma Points

! Compute a set of sigma points ! How to choose the sigma points?
! Each sigma points has a weight ! How to set the weights?
! Transform the point through the non-
linear function
! Compute a Gaussian from weighted
points

## ! Avoids to linearize around the

mean as the EKF does
7 8
Sigma Points Properties Sigma Points
! How to choose the sigma points? ! Choosing the sigma points
! How to set the weights?
! Select so that:

9 10

## Sigma Points Matrix Square Root

! Choosing the sigma points ! Defined as
! Computed via diagonalization

root

## dimensionality scaling parameter

11 12
Matrix Square Root Cholesky Matrix Square Root
! Thus, we can define ! Alternative definition of the matrix
square root

## ! Result of the Cholesky decomposition

! Numerically stable solution
! so that
! Often used in UKF implementations
! L and have the same Eigenvectors

13 14

## Sigma Points and Eigenvectors Sigma Points Example

! Sigma point can but do not have to
lie on the main axes of

15 16
Sigma Point Weights Recover the Gaussian
! Weight sigma points ! Compute Gaussian from weighted and
transformed points
for computing parameters
the mean

## for computing the covariance

17 18

Example Examples

19 20
Unscented Transform Summary UT Parameters
! Sigma points ! Free parameters as there is no unique
solution
! Scales Unscented Transform uses

## Influence how far the

! Weights sigma points are
away from the mean

## Optimal choice for

Gaussians
21 22

Examples Examples

23 24
EKF Algorithm EKF to UKF Prediction
Unscented

## replace this by sigma point

propagation of the motion

25 26

Unscented

## replace this by sigma point

propagation of the motion

## use sigma point propagation for the

expected observation and Kalman gain

27 28
UKF Algorithm Correction (1) UKF Algorithm Correction (1)

(from EKF)
29 30

## UKF Algorithm Correction (2) UKF Algorithm Correction (2)

31
(see next slide) 32
From EKF to UKF Computing UKF vs. EKF
the Covariance

33 34

UKF vs. EKF (Small Covariance) UKF vs. EKF Banana Shape

EKF approximation

UKF approximation

35 36
UKF vs. EKF UT/UKF Summary
! Unscented transforms as an
alternative to linearization
! UT is a better approximation than
Taylor expansion
! UT uses sigma point propagation
! Free parameters in UT
! UKF uses the UT in the prediction and
correction step

## UKF vs. EKF Literature

! Same results as EKF for linear models Unscented Transform and UKF
! Better approximation than EKF for ! Thrun et al.: Probabilistic Robotics,
non-linear models Chapter 3.4
! Differences often somewhat small ! A New Extension of the Kalman Filter
! No Jacobians needed for the UKF to Nonlinear Systems by Julier and
! Same complexity class Uhlmann, 1995
! Dynamische Zustandsschtzung by
! Slightly slower than the EKF
Frnken, 2006, pages 31-34
! Still requires Gaussian distributions

39 40