Lecture 18: Particle Filters

Wei Du

INFO0013 - Computer Vision

Du Particle Filters: Theory and Applications

Objective

from an Engineer’s Point of View

an overview of Particle Filters

applications on visual tracking

Du Particle Filters: Theory and Applications

Introduction

In general, Particle Filter is an extension to Kalman Filter.

it deals with the same problems that Kalman Filter does
but aims at different issues:
Non-Gaussianity
Non-Linearity

The advantages:
easy to implement
a general solution to a large range of problems

Du Particle Filters: Theory and Applications

History of Particle Filters

It was started by physicists and mathematicians in Project

Manhattan, but was forgotten soon.

Kalman filter was proposed in 1960.

It was remembered in 70s and developed in 80s.

It was first introduced into visual tracking by Michael Isard -

“Condensation”, the best paper of ECCV1996.

In 1996 - 2000, Oxford Heritage

Blossom in 2001 - 2002

Du Particle Filters: Theory and Applications

An Abstract View of Tracking
The entity being tracked is a state vector x that cannot be
observed directly.
We have a dynamic model P(xt |xt−1 ) of the state that
predicts how it evolves over time.
Periodically, we obtain an observation vector Y that is a
function of the state.
Using the observation P(yt |xt ), we correct our internal
estimate of the state.

Du Particle Filters: Theory and Applications

Problem Statement
An estimate of xt based on all the measurements y1:t
Z
x̂t = E[xt |y1:t ] = xt P(xt |y1:t ) d xt
xt

How to infer the posterior pdf

P(xt |y1:t )

Du Particle Filters: Theory and Applications

Inference Recursion

Independence assumptions:
Only the immediate past matters(Markov assumption):

P(xt |x1:t−1 ) = P(xt |xt−1 )

Measurements depend only on the current state:

P(yt |x1:t ) = P(yt |xt )

Recursion (Bayes’s law)

P(yt |xt ) P(xt |y1:t−1 )
P(xt |y1:t ) = P(yt ) ∝ P(yt |xt ) P(xt |y1:t−1 )
Z
∝ P(yt |xt ) P(xt |xt−1 ) P(xt−1 |y1:t−1 ) d xt−1
xt−1

Du Particle Filters: Theory and Applications

Inference Recursion

Z
P(xt |y1:t ) ∝ P(yt |xt ) P(xt |xt−1 ) P(xt−1 |y1:t−1 ) d xt−1
xt−1

Z
predict: P(xt |y1:t−1 ) = P(xt |xt−1 ) P(xt−1 |y1:t−1 ) d xt−1
xt−1

correct: P(xt |y1:t ) ∝ P(yt |xt ) P(xt |y1:t−1 )

Du Particle Filters: Theory and Applications

Kalman Filter

linear motion and Gaussian distribution

Z
predict: P(xt |y1:t−1 ) = P(xt |xt−1 ) P(xt−1 |y1:t−1 ) d xt−1
xt−1

correct: P(xt |y1:t ) ∝ P(yt |xt ) P(xt |y1:t−1 )

Du Particle Filters: Theory and Applications

Extended Kalman Filter

non-linear motion and Gaussian distribution

Z
predict: P(xt |y1:t−1 ) = P(xt |xt−1 ) P(xt−1 |y1:t−1 ) d xt−1
xt−1

correct: P(xt |y1:t ) ∝ P(yt |xt ) P(xt |y1:t−1 )

Du Particle Filters: Theory and Applications

Particle Filter

linear or non-linear motion and non-Gaussian distribution

Z
predict: P(xt |y1:t−1 ) = P(xt |xt−1 ) P(xt−1 |y1:t−1 ) d xt−1
xt−1

correct: P(xt |y1:t ) ∝ P(yt |xt ) P(xt |y1:t−1 )

Du Particle Filters: Theory and Applications

The representation of arbitrary distributions

Du Particle Filters: Theory and Applications

The representation of arbitrary distributions

mean+covariance for Gaussian distribution

fail to capture real-world densities.

Du Particle Filters: Theory and Applications

The representation of arbitrary distributions

mixture of Gaussian

other kernel based representations

Mixture models are appealing, but hard to propagate

analytically.
Du Particle Filters: Theory and Applications
The representation of arbitrary distributions

histogram

Du Particle Filters: Theory and Applications

The representation of arbitrary distributions

weighted samples/particles (sti , πti ), πti ∝ p(sti |y1:t )

Du Particle Filters: Theory and Applications

What is Particle Filter?

Particle filters propagate and maintain the posterior pdf of the

target states in time,
p(xt |y1:t )
represented by a set of weighted particles,

(sti , πti )

Condensation – CONditional DENSity propagATION (The first

paper of particle filters in computer vision)

Du Particle Filters: Theory and Applications

Monte Carlo Method
Based on random experiments or trials
Eg. Choose points randomly inside the square
The area under the curve f (x) is
Z b
number of blue points
= f (x) d x ≈ × area of the square
a total number of points
As the total number of random points increase, the
accuracy gets better

Du Particle Filters: Theory and Applications

Factored Sampling

p(xt |y1:t ) ∼ (sti , πti )

An estimate of xt based on all the measurements y1:t

Z X
x̂t = E[xt |y1:t ] = xt P(xt |y1:t ) d xt ≈ sti πti
xt i

Du Particle Filters: Theory and Applications

Overview of Particle Filter Tracking

Du Particle Filters: Theory and Applications

Objective

Du Particle Filters: Theory and Applications

Step 1: Sample

Du Particle Filters: Theory and Applications

Step 2: Predict

Du Particle Filters: Theory and Applications

Step 3: Weight

Du Particle Filters: Theory and Applications

Step 4: Normalize

Du Particle Filters: Theory and Applications

The algorithm

Initialize: Draw N particles s0i ∼ P(x0 ), π0i = 1/N

For t = 1 to end
i
Resample s̃t−1 i
∼ (st−1 i
, πt−1 i
) so that π̃t−1 = 1/N

Predict sti ∼ P(sti |s̃t−1

i i
) = N(f (st−1 ); Σt )
Weight πti ∼ P(yt |sti ) for each particle sti .
Estimate x̂t = sti πti
End For

Du Particle Filters: Theory and Applications

Color-based Particle Filter Tracking

Du Particle Filters: Theory and Applications

Importance Sampling

Draw particles from a proposal function q(sti |yt )

P(sti |yt )
Assign weights πti ∼ q(sti |yt )
to the particles sti

Du Particle Filters: Theory and Applications

Demo

Du Particle Filters: Theory and Applications

Is Particle Filter always better than Kalman Filter?

Particle Filter is an approximation, and the precision depends

on the number of particles sampled.

So it should be the last choice when you don’t have the

analytical representations.

Du Particle Filters: Theory and Applications

How many particles do you need to approximate the
posterior?

Impossible to have a rule to say how many are enough.

The number increases exponentially with the dimension of the

state space.

The more, the better. So sample as many as you can afford.

Du Particle Filters: Theory and Applications

Conclusions

Particle filters are cool.

40 lines in Matlab can do a basic particle filter.

To make your tracker work, 20000 extra lines are required.

Du Particle Filters: Theory and Applications

Particle Filter Recourses

Sequential Monte Carlo Methods Homepage:

http://www-sigproc.eng.cam.ac.uk/smc/index.html

Condensation Homepage

http://www.robots.ox.ac.uk/ misard/condensation.html

Jean-Bernard Hayet’s Homepage

http://www.montefiore.ulg.ac.be/ jbhayet/ltilib.php

Bible of Particle Filters

Sequential Monte Carlo Methods in Practice

Du Particle Filters: Theory and Applications