the particle filtering methodology in signal · pdf filethe particle filtering methodology in...
TRANSCRIPT
![Page 1: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/1.jpg)
The Particle Filtering Methodology in Signal Processing
Petar M. Djuric
Department of Electrical & Computer EngineeringStony Brook University
September 25, 2009
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 1 / 70
![Page 2: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/2.jpg)
Outline
Preface
Many problems in adaptive filtering are nonlinear and non-Gaussian
Of the many methods that have been proposed in the literature forsolving such problems, particle filtering (PF) has become one of themost popular
In this talk the basics of PF are revisited and a review of some of itsmost important implementations is discussed
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 2 / 70
![Page 3: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/3.jpg)
Outline
Preface
Many problems in adaptive filtering are nonlinear and non-Gaussian
Of the many methods that have been proposed in the literature forsolving such problems, particle filtering (PF) has become one of themost popular
In this talk the basics of PF are revisited and a review of some of itsmost important implementations is discussed
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 2 / 70
![Page 4: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/4.jpg)
Outline
Preface
Many problems in adaptive filtering are nonlinear and non-Gaussian
Of the many methods that have been proposed in the literature forsolving such problems, particle filtering (PF) has become one of themost popular
In this talk the basics of PF are revisited and a review of some of itsmost important implementations is discussed
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 2 / 70
![Page 5: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/5.jpg)
Outline
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 3 / 70
![Page 6: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/6.jpg)
Outline
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 3 / 70
![Page 7: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/7.jpg)
Outline
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 3 / 70
![Page 8: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/8.jpg)
Outline
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 3 / 70
![Page 9: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/9.jpg)
Outline
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 3 / 70
![Page 10: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/10.jpg)
Outline
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 3 / 70
![Page 11: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/11.jpg)
Outline
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 3 / 70
![Page 12: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/12.jpg)
Outline
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 3 / 70
![Page 13: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/13.jpg)
Outline
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 3 / 70
![Page 14: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/14.jpg)
Outline
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 3 / 70
![Page 15: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/15.jpg)
Introduction
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 4 / 70
![Page 16: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/16.jpg)
Introduction
Introducing PF
Some of the most important problems in signal processing requiresequential processing of data
The data describe a system that is mathematically represented byequations, which model the system’s “random” evolution with time
The system is defined by its state variables of which some aredynamic and some are constant
A typical assumption about the state is that it is Markovian
The state is not directly observable and we acquire measurementswhich are degraded by noise and obtained sequentially in time
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 5 / 70
![Page 17: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/17.jpg)
Introduction
Introducing PF
Some of the most important problems in signal processing requiresequential processing of data
The data describe a system that is mathematically represented byequations, which model the system’s “random” evolution with time
The system is defined by its state variables of which some aredynamic and some are constant
A typical assumption about the state is that it is Markovian
The state is not directly observable and we acquire measurementswhich are degraded by noise and obtained sequentially in time
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 5 / 70
![Page 18: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/18.jpg)
Introduction
Introducing PF
Some of the most important problems in signal processing requiresequential processing of data
The data describe a system that is mathematically represented byequations, which model the system’s “random” evolution with time
The system is defined by its state variables of which some aredynamic and some are constant
A typical assumption about the state is that it is Markovian
The state is not directly observable and we acquire measurementswhich are degraded by noise and obtained sequentially in time
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 5 / 70
![Page 19: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/19.jpg)
Introduction
Introducing PF
Some of the most important problems in signal processing requiresequential processing of data
The data describe a system that is mathematically represented byequations, which model the system’s “random” evolution with time
The system is defined by its state variables of which some aredynamic and some are constant
A typical assumption about the state is that it is Markovian
The state is not directly observable and we acquire measurementswhich are degraded by noise and obtained sequentially in time
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 5 / 70
![Page 20: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/20.jpg)
Introduction
Introducing PF
Some of the most important problems in signal processing requiresequential processing of data
The data describe a system that is mathematically represented byequations, which model the system’s “random” evolution with time
The system is defined by its state variables of which some aredynamic and some are constant
A typical assumption about the state is that it is Markovian
The state is not directly observable and we acquire measurementswhich are degraded by noise and obtained sequentially in time
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 5 / 70
![Page 21: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/21.jpg)
Introduction
Goal
The main objective of sequential signal processing is the recursiveestimation of the state of the system based on the availablemeasurements
0 20 40 60 80 100 120 140 160 180 200−15
−10
−5
0
5
10
t (sec)
x t
0 20 40 60 80 100 120 140 160 180 200−15
−10
−5
0
5
10
15
t (sec)y t
The methods that are developed for estimation of the state are calledfilters
Estimates of the state in the past or prediction of its values in thefuture may also be of interest
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 6 / 70
![Page 22: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/22.jpg)
Introduction
Goal
The main objective of sequential signal processing is the recursiveestimation of the state of the system based on the availablemeasurements
0 20 40 60 80 100 120 140 160 180 200−15
−10
−5
0
5
10
t (sec)
x t
0 20 40 60 80 100 120 140 160 180 200−15
−10
−5
0
5
10
15
t (sec)y t
The methods that are developed for estimation of the state are calledfilters
Estimates of the state in the past or prediction of its values in thefuture may also be of interest
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 6 / 70
![Page 23: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/23.jpg)
Introduction
Goal
The main objective of sequential signal processing is the recursiveestimation of the state of the system based on the availablemeasurements
0 20 40 60 80 100 120 140 160 180 200−15
−10
−5
0
5
10
t (sec)
x t
0 20 40 60 80 100 120 140 160 180 200−15
−10
−5
0
5
10
15
t (sec)y t
The methods that are developed for estimation of the state are calledfilters
Estimates of the state in the past or prediction of its values in thefuture may also be of interest
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 6 / 70
![Page 24: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/24.jpg)
Introduction
Optimal filtering methods
Optimal filtering methods for sequential signal processing are basedon analytical updates of the densities of interest
↓
Integration is a key operation
Closed form solutions are possible in a very small number of situations
Gaussian noiseand
linear functions
→ Optimal estimation with Kalman filtering
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 7 / 70
![Page 25: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/25.jpg)
Introduction
Optimal filtering methods
Optimal filtering methods for sequential signal processing are basedon analytical updates of the densities of interest
↓
Integration is a key operation
Closed form solutions are possible in a very small number of situations
Gaussian noiseand
linear functions
→ Optimal estimation with Kalman filtering
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 7 / 70
![Page 26: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/26.jpg)
Introduction
Optimal filtering methods
Optimal filtering methods for sequential signal processing are basedon analytical updates of the densities of interest
↓
Integration is a key operation
Closed form solutions are possible in a very small number of situations
Gaussian noiseand
linear functions
→ Optimal estimation with Kalman filtering
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 7 / 70
![Page 27: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/27.jpg)
Introduction
Suboptimal filtering methods
Non-Gaussian noiseor
nonlinear functions
→ Numerical techniques
Extended Kalman filter [Anderson & Moore, 1979]
Gaussian sum filters [Alspach & Sorenson, 1972]
Methods based on approximations of the first two moments of thedensities [Masreliez, 1975; West, 1985]
Evaluation of densities over grids [Kitagawa, 1987; Sorenson, 1988]
The unscented Kalman filter [Julier et al., 2000]
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 8 / 70
![Page 28: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/28.jpg)
Introduction
Suboptimal filtering methods
Non-Gaussian noiseor
nonlinear functions
→ Numerical techniques
Extended Kalman filter [Anderson & Moore, 1979]
Gaussian sum filters [Alspach & Sorenson, 1972]
Methods based on approximations of the first two moments of thedensities [Masreliez, 1975; West, 1985]
Evaluation of densities over grids [Kitagawa, 1987; Sorenson, 1988]
The unscented Kalman filter [Julier et al., 2000]
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 8 / 70
![Page 29: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/29.jpg)
Introduction
Suboptimal filtering methods
Non-Gaussian noiseor
nonlinear functions
→ Numerical techniques
Extended Kalman filter [Anderson & Moore, 1979]
Gaussian sum filters [Alspach & Sorenson, 1972]
Methods based on approximations of the first two moments of thedensities [Masreliez, 1975; West, 1985]
Evaluation of densities over grids [Kitagawa, 1987; Sorenson, 1988]
The unscented Kalman filter [Julier et al., 2000]
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 8 / 70
![Page 30: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/30.jpg)
Introduction
Suboptimal filtering methods
Non-Gaussian noiseor
nonlinear functions
→ Numerical techniques
Extended Kalman filter [Anderson & Moore, 1979]
Gaussian sum filters [Alspach & Sorenson, 1972]
Methods based on approximations of the first two moments of thedensities [Masreliez, 1975; West, 1985]
Evaluation of densities over grids [Kitagawa, 1987; Sorenson, 1988]
The unscented Kalman filter [Julier et al., 2000]
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 8 / 70
![Page 31: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/31.jpg)
Introduction
Suboptimal filtering methods
Non-Gaussian noiseor
nonlinear functions
→ Numerical techniques
Extended Kalman filter [Anderson & Moore, 1979]
Gaussian sum filters [Alspach & Sorenson, 1972]
Methods based on approximations of the first two moments of thedensities [Masreliez, 1975; West, 1985]
Evaluation of densities over grids [Kitagawa, 1987; Sorenson, 1988]
The unscented Kalman filter [Julier et al., 2000]
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 8 / 70
![Page 32: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/32.jpg)
Introduction
Suboptimal filtering methods
Non-Gaussian noiseor
nonlinear functions
→ Numerical techniques
Extended Kalman filter [Anderson & Moore, 1979]
Gaussian sum filters [Alspach & Sorenson, 1972]
Methods based on approximations of the first two moments of thedensities [Masreliez, 1975; West, 1985]
Evaluation of densities over grids [Kitagawa, 1987; Sorenson, 1988]
The unscented Kalman filter [Julier et al., 2000]
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 8 / 70
![Page 33: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/33.jpg)
Introduction
PF methodology
PF is very similar in spirit as the one that exploits deterministic grids
However, PF uses random grids, which means that the location of thenodes vary with time in a random way
At one time instant the grid is composed of one set of nodes, and atthe next time instant, this set of nodes is completely different
The nodes are called particles, and they have assigned weights, whichcan be interpreted as probability masses
The particles and the weights form a discrete random measure, whichis used to approximate the densities of interest
In the generation of particles, each particle has a “parent”, and theparent its own parent and so on
Such sequence of particles is called a particle stream
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 9 / 70
![Page 34: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/34.jpg)
Introduction
PF methodology
PF is very similar in spirit as the one that exploits deterministic grids
However, PF uses random grids, which means that the location of thenodes vary with time in a random way
At one time instant the grid is composed of one set of nodes, and atthe next time instant, this set of nodes is completely different
The nodes are called particles, and they have assigned weights, whichcan be interpreted as probability masses
The particles and the weights form a discrete random measure, whichis used to approximate the densities of interest
In the generation of particles, each particle has a “parent”, and theparent its own parent and so on
Such sequence of particles is called a particle stream
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 9 / 70
![Page 35: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/35.jpg)
Introduction
PF methodology
PF is very similar in spirit as the one that exploits deterministic grids
However, PF uses random grids, which means that the location of thenodes vary with time in a random way
At one time instant the grid is composed of one set of nodes, and atthe next time instant, this set of nodes is completely different
The nodes are called particles, and they have assigned weights, whichcan be interpreted as probability masses
The particles and the weights form a discrete random measure, whichis used to approximate the densities of interest
In the generation of particles, each particle has a “parent”, and theparent its own parent and so on
Such sequence of particles is called a particle stream
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 9 / 70
![Page 36: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/36.jpg)
Introduction
PF methodology
PF is very similar in spirit as the one that exploits deterministic grids
However, PF uses random grids, which means that the location of thenodes vary with time in a random way
At one time instant the grid is composed of one set of nodes, and atthe next time instant, this set of nodes is completely different
The nodes are called particles, and they have assigned weights, whichcan be interpreted as probability masses
The particles and the weights form a discrete random measure, whichis used to approximate the densities of interest
In the generation of particles, each particle has a “parent”, and theparent its own parent and so on
Such sequence of particles is called a particle stream
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 9 / 70
![Page 37: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/37.jpg)
Introduction
PF methodology
PF is very similar in spirit as the one that exploits deterministic grids
However, PF uses random grids, which means that the location of thenodes vary with time in a random way
At one time instant the grid is composed of one set of nodes, and atthe next time instant, this set of nodes is completely different
The nodes are called particles, and they have assigned weights, whichcan be interpreted as probability masses
The particles and the weights form a discrete random measure, whichis used to approximate the densities of interest
In the generation of particles, each particle has a “parent”, and theparent its own parent and so on
Such sequence of particles is called a particle stream
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 9 / 70
![Page 38: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/38.jpg)
Introduction
PF methodology
PF is very similar in spirit as the one that exploits deterministic grids
However, PF uses random grids, which means that the location of thenodes vary with time in a random way
At one time instant the grid is composed of one set of nodes, and atthe next time instant, this set of nodes is completely different
The nodes are called particles, and they have assigned weights, whichcan be interpreted as probability masses
The particles and the weights form a discrete random measure, whichis used to approximate the densities of interest
In the generation of particles, each particle has a “parent”, and theparent its own parent and so on
Such sequence of particles is called a particle stream
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 9 / 70
![Page 39: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/39.jpg)
Introduction
PF methodology
PF is very similar in spirit as the one that exploits deterministic grids
However, PF uses random grids, which means that the location of thenodes vary with time in a random way
At one time instant the grid is composed of one set of nodes, and atthe next time instant, this set of nodes is completely different
The nodes are called particles, and they have assigned weights, whichcan be interpreted as probability masses
The particles and the weights form a discrete random measure, whichis used to approximate the densities of interest
In the generation of particles, each particle has a “parent”, and theparent its own parent and so on
Such sequence of particles is called a particle stream
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 9 / 70
![Page 40: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/40.jpg)
Introduction
PF philosophy
A challenge in implementing PF is the placement of the nodes of thegrids
The generation of particles should be in regions of the state spaceover which the densities carry significant probability masses, avoidingparts of the state space with negligible probability masses
To that end, one exploits concepts from statistics known asimportance sampling (IS) and sampling-importance-resampling (SIR)
Bayesian theory is applied to the computation of the particle weights
The sequential processing amounts to recursive updating of thediscrete random measure with the arrival of new observations, wherethe updates correspond to generation of new particles andcomputation of their weights
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 10 / 70
![Page 41: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/41.jpg)
Introduction
PF philosophy
A challenge in implementing PF is the placement of the nodes of thegrids
The generation of particles should be in regions of the state spaceover which the densities carry significant probability masses, avoidingparts of the state space with negligible probability masses
To that end, one exploits concepts from statistics known asimportance sampling (IS) and sampling-importance-resampling (SIR)
Bayesian theory is applied to the computation of the particle weights
The sequential processing amounts to recursive updating of thediscrete random measure with the arrival of new observations, wherethe updates correspond to generation of new particles andcomputation of their weights
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 10 / 70
![Page 42: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/42.jpg)
Introduction
PF philosophy
A challenge in implementing PF is the placement of the nodes of thegrids
The generation of particles should be in regions of the state spaceover which the densities carry significant probability masses, avoidingparts of the state space with negligible probability masses
To that end, one exploits concepts from statistics known asimportance sampling (IS) and sampling-importance-resampling (SIR)
Bayesian theory is applied to the computation of the particle weights
The sequential processing amounts to recursive updating of thediscrete random measure with the arrival of new observations, wherethe updates correspond to generation of new particles andcomputation of their weights
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 10 / 70
![Page 43: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/43.jpg)
Introduction
PF philosophy
A challenge in implementing PF is the placement of the nodes of thegrids
The generation of particles should be in regions of the state spaceover which the densities carry significant probability masses, avoidingparts of the state space with negligible probability masses
To that end, one exploits concepts from statistics known asimportance sampling (IS) and sampling-importance-resampling (SIR)
Bayesian theory is applied to the computation of the particle weights
The sequential processing amounts to recursive updating of thediscrete random measure with the arrival of new observations, wherethe updates correspond to generation of new particles andcomputation of their weights
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 10 / 70
![Page 44: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/44.jpg)
Introduction
PF philosophy
A challenge in implementing PF is the placement of the nodes of thegrids
The generation of particles should be in regions of the state spaceover which the densities carry significant probability masses, avoidingparts of the state space with negligible probability masses
To that end, one exploits concepts from statistics known asimportance sampling (IS) and sampling-importance-resampling (SIR)
Bayesian theory is applied to the computation of the particle weights
The sequential processing amounts to recursive updating of thediscrete random measure with the arrival of new observations, wherethe updates correspond to generation of new particles andcomputation of their weights
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 10 / 70
![Page 45: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/45.jpg)
Introduction
Some history
The roots of PF were established about fifty years ago with methodsfor estimating the mean squared extension of a self-avoiding randomwalk on lattice spaces
One of the first applications of the methodology was on thesimulation of chain polymers
The control community produced interesting work on sequentialMonte Carlo integration methods in the sixties and seventies
The popularity of PF in the last fifteen years was triggered by theapplication of the SIR filter to tracking problems
The timing (1993) was perfect because it came during a period whencomputing power started to become widely available
Ever since, the amount of work on particle filtering has proliferatedand many important advances have been made
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 11 / 70
![Page 46: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/46.jpg)
Introduction
Some history
The roots of PF were established about fifty years ago with methodsfor estimating the mean squared extension of a self-avoiding randomwalk on lattice spaces
One of the first applications of the methodology was on thesimulation of chain polymers
The control community produced interesting work on sequentialMonte Carlo integration methods in the sixties and seventies
The popularity of PF in the last fifteen years was triggered by theapplication of the SIR filter to tracking problems
The timing (1993) was perfect because it came during a period whencomputing power started to become widely available
Ever since, the amount of work on particle filtering has proliferatedand many important advances have been made
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 11 / 70
![Page 47: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/47.jpg)
Introduction
Some history
The roots of PF were established about fifty years ago with methodsfor estimating the mean squared extension of a self-avoiding randomwalk on lattice spaces
One of the first applications of the methodology was on thesimulation of chain polymers
The control community produced interesting work on sequentialMonte Carlo integration methods in the sixties and seventies
The popularity of PF in the last fifteen years was triggered by theapplication of the SIR filter to tracking problems
The timing (1993) was perfect because it came during a period whencomputing power started to become widely available
Ever since, the amount of work on particle filtering has proliferatedand many important advances have been made
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 11 / 70
![Page 48: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/48.jpg)
Introduction
Some history
The roots of PF were established about fifty years ago with methodsfor estimating the mean squared extension of a self-avoiding randomwalk on lattice spaces
One of the first applications of the methodology was on thesimulation of chain polymers
The control community produced interesting work on sequentialMonte Carlo integration methods in the sixties and seventies
The popularity of PF in the last fifteen years was triggered by theapplication of the SIR filter to tracking problems
The timing (1993) was perfect because it came during a period whencomputing power started to become widely available
Ever since, the amount of work on particle filtering has proliferatedand many important advances have been made
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 11 / 70
![Page 49: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/49.jpg)
Introduction
Some history
The roots of PF were established about fifty years ago with methodsfor estimating the mean squared extension of a self-avoiding randomwalk on lattice spaces
One of the first applications of the methodology was on thesimulation of chain polymers
The control community produced interesting work on sequentialMonte Carlo integration methods in the sixties and seventies
The popularity of PF in the last fifteen years was triggered by theapplication of the SIR filter to tracking problems
The timing (1993) was perfect because it came during a period whencomputing power started to become widely available
Ever since, the amount of work on particle filtering has proliferatedand many important advances have been made
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 11 / 70
![Page 50: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/50.jpg)
Introduction
Some history
The roots of PF were established about fifty years ago with methodsfor estimating the mean squared extension of a self-avoiding randomwalk on lattice spaces
One of the first applications of the methodology was on thesimulation of chain polymers
The control community produced interesting work on sequentialMonte Carlo integration methods in the sixties and seventies
The popularity of PF in the last fifteen years was triggered by theapplication of the SIR filter to tracking problems
The timing (1993) was perfect because it came during a period whencomputing power started to become widely available
Ever since, the amount of work on particle filtering has proliferatedand many important advances have been made
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 11 / 70
![Page 51: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/51.jpg)
Introduction
Why PF?
One driving force for the advancement of PF is the ever increasingrange of its applications
PF can be applied to any state space model where the likelihood andthe prior are computable up to proportionality constants
Its accuracy depends on how well we generate the particles and howmany particles we use to represent the random measure
It is well known that PF is computationally intensive which is an issuebecause in sequential signal processing the required operations mustbe completed before a deadline
However, PF allows for significant parallelization of its operations
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 12 / 70
![Page 52: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/52.jpg)
Introduction
Why PF?
One driving force for the advancement of PF is the ever increasingrange of its applications
PF can be applied to any state space model where the likelihood andthe prior are computable up to proportionality constants
Its accuracy depends on how well we generate the particles and howmany particles we use to represent the random measure
It is well known that PF is computationally intensive which is an issuebecause in sequential signal processing the required operations mustbe completed before a deadline
However, PF allows for significant parallelization of its operations
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 12 / 70
![Page 53: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/53.jpg)
Introduction
Why PF?
One driving force for the advancement of PF is the ever increasingrange of its applications
PF can be applied to any state space model where the likelihood andthe prior are computable up to proportionality constants
Its accuracy depends on how well we generate the particles and howmany particles we use to represent the random measure
It is well known that PF is computationally intensive which is an issuebecause in sequential signal processing the required operations mustbe completed before a deadline
However, PF allows for significant parallelization of its operations
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 12 / 70
![Page 54: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/54.jpg)
Introduction
Why PF?
One driving force for the advancement of PF is the ever increasingrange of its applications
PF can be applied to any state space model where the likelihood andthe prior are computable up to proportionality constants
Its accuracy depends on how well we generate the particles and howmany particles we use to represent the random measure
It is well known that PF is computationally intensive which is an issuebecause in sequential signal processing the required operations mustbe completed before a deadline
However, PF allows for significant parallelization of its operations
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 12 / 70
![Page 55: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/55.jpg)
Introduction
Why PF?
One driving force for the advancement of PF is the ever increasingrange of its applications
PF can be applied to any state space model where the likelihood andthe prior are computable up to proportionality constants
Its accuracy depends on how well we generate the particles and howmany particles we use to represent the random measure
It is well known that PF is computationally intensive which is an issuebecause in sequential signal processing the required operations mustbe completed before a deadline
However, PF allows for significant parallelization of its operations
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 12 / 70
![Page 56: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/56.jpg)
Introduction
Applications of PF
Target tracking, positioning and navigation
Robotics
Communications (detection in flat fading, equalization andsynchronization)
Speech processing for speech enhancement
Seismic signal processing for blind deconvolution
System engineering for fault detection
Computer vision
Econometrics
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 13 / 70
![Page 57: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/57.jpg)
Introduction
Applications of PF
Target tracking, positioning and navigation
Robotics
Communications (detection in flat fading, equalization andsynchronization)
Speech processing for speech enhancement
Seismic signal processing for blind deconvolution
System engineering for fault detection
Computer vision
Econometrics
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 13 / 70
![Page 58: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/58.jpg)
Introduction
Applications of PF
Target tracking, positioning and navigation
Robotics
Communications (detection in flat fading, equalization andsynchronization)
Speech processing for speech enhancement
Seismic signal processing for blind deconvolution
System engineering for fault detection
Computer vision
Econometrics
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 13 / 70
![Page 59: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/59.jpg)
Introduction
Applications of PF
Target tracking, positioning and navigation
Robotics
Communications (detection in flat fading, equalization andsynchronization)
Speech processing for speech enhancement
Seismic signal processing for blind deconvolution
System engineering for fault detection
Computer vision
Econometrics
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 13 / 70
![Page 60: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/60.jpg)
Introduction
Applications of PF
Target tracking, positioning and navigation
Robotics
Communications (detection in flat fading, equalization andsynchronization)
Speech processing for speech enhancement
Seismic signal processing for blind deconvolution
System engineering for fault detection
Computer vision
Econometrics
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 13 / 70
![Page 61: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/61.jpg)
Introduction
Applications of PF
Target tracking, positioning and navigation
Robotics
Communications (detection in flat fading, equalization andsynchronization)
Speech processing for speech enhancement
Seismic signal processing for blind deconvolution
System engineering for fault detection
Computer vision
Econometrics
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 13 / 70
![Page 62: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/62.jpg)
Introduction
Applications of PF
Target tracking, positioning and navigation
Robotics
Communications (detection in flat fading, equalization andsynchronization)
Speech processing for speech enhancement
Seismic signal processing for blind deconvolution
System engineering for fault detection
Computer vision
Econometrics
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 13 / 70
![Page 63: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/63.jpg)
Introduction
Applications of PF
Target tracking, positioning and navigation
Robotics
Communications (detection in flat fading, equalization andsynchronization)
Speech processing for speech enhancement
Seismic signal processing for blind deconvolution
System engineering for fault detection
Computer vision
Econometrics
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 13 / 70
![Page 64: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/64.jpg)
Motivation for use of particle filtering
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 14 / 70
![Page 65: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/65.jpg)
Motivation for use of particle filtering
Representation of dynamic systems
Many problems can be stated in terms of dynamic systems
1 State equation: hidden random signal to be estimated
xt = f (xt−1,ut)
xt signal (state) of interestf state transition function (possibly nonlinear)ut state perturbation noise
2 Observation equation: available information
yt = g (xt , vt)
yt observed signalg measurement function (possibly nonlinear)vt observation noise
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 15 / 70
![Page 66: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/66.jpg)
Motivation for use of particle filtering
Representation of dynamic systems
Many problems can be stated in terms of dynamic systems
1 State equation: hidden random signal to be estimated
xt = f (xt−1,ut)
xt signal (state) of interestf state transition function (possibly nonlinear)ut state perturbation noise
2 Observation equation: available information
yt = g (xt , vt)
yt observed signalg measurement function (possibly nonlinear)vt observation noise
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 15 / 70
![Page 67: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/67.jpg)
Motivation for use of particle filtering
Alternative representation of dynamic systems
One can also represent the dynamic system in terms of densities
1 State equation: p (xt |xt−1,θ)
2 Observation equation: p (yt |xt ,ψ)
The form of the density functions depends on:
the functions f (·) and g(·)the densities of ut and vt
θ ∈ <Lθ
ψ ∈ <Lψ
}→ Unknown fixed parameter vectors
They can be included in xt as static parameters
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 16 / 70
![Page 68: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/68.jpg)
Motivation for use of particle filtering
Important densities
Three densities play a critical role in sequential signal processing
1 Filtering density: p (xt |y1:t), where y1:t = {y1, y2, · · · , yt}
2 Predictive density: p (xt+l |y1:t), where l ≥ 1
3 Smoothing density: p (xt |y1:T ), where T > t andy1:T = {y1, y2, · · · , yT}
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 17 / 70
![Page 69: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/69.jpg)
Motivation for use of particle filtering
Important densities
Three densities play a critical role in sequential signal processing
1 Filtering density: p (xt |y1:t), where y1:t = {y1, y2, · · · , yt}
2 Predictive density: p (xt+l |y1:t), where l ≥ 1
3 Smoothing density: p (xt |y1:T ), where T > t andy1:T = {y1, y2, · · · , yT}
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 17 / 70
![Page 70: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/70.jpg)
Motivation for use of particle filtering
Important densities
Three densities play a critical role in sequential signal processing
1 Filtering density: p (xt |y1:t), where y1:t = {y1, y2, · · · , yt}
2 Predictive density: p (xt+l |y1:t), where l ≥ 1
3 Smoothing density: p (xt |y1:T ), where T > t andy1:T = {y1, y2, · · · , yT}
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 17 / 70
![Page 71: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/71.jpg)
Motivation for use of particle filtering
Tracking the important densities
Objective: Track the densities by exploiting recursive relationships
1 Filtering density: p (xt |y1:t) from p(xt−1|y1:t−1
)2 Predictive density: p (xt+l |y1:t) from p (xt+l−1|y1:t)
3 Smoothing density: p (xt |y1:T ) from p (xt+1|y1:T )
Complete information about xt is given by these densities
An algorithm that can track these densities exactly is called anoptimal algorithm
In many practical situations, optimal algorithms are impossible toimplement
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 18 / 70
![Page 72: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/72.jpg)
Motivation for use of particle filtering
Tracking the important densities
Objective: Track the densities by exploiting recursive relationships
1 Filtering density: p (xt |y1:t) from p(xt−1|y1:t−1
)2 Predictive density: p (xt+l |y1:t) from p (xt+l−1|y1:t)
3 Smoothing density: p (xt |y1:T ) from p (xt+1|y1:T )
Complete information about xt is given by these densities
An algorithm that can track these densities exactly is called anoptimal algorithm
In many practical situations, optimal algorithms are impossible toimplement
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 18 / 70
![Page 73: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/73.jpg)
Motivation for use of particle filtering
Tracking the important densities
Objective: Track the densities by exploiting recursive relationships
1 Filtering density: p (xt |y1:t) from p(xt−1|y1:t−1
)2 Predictive density: p (xt+l |y1:t) from p (xt+l−1|y1:t)
3 Smoothing density: p (xt |y1:T ) from p (xt+1|y1:T )
Complete information about xt is given by these densities
An algorithm that can track these densities exactly is called anoptimal algorithm
In many practical situations, optimal algorithms are impossible toimplement
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 18 / 70
![Page 74: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/74.jpg)
Motivation for use of particle filtering
Tracking the important densities
Objective: Track the densities by exploiting recursive relationships
1 Filtering density: p (xt |y1:t) from p(xt−1|y1:t−1
)2 Predictive density: p (xt+l |y1:t) from p (xt+l−1|y1:t)
3 Smoothing density: p (xt |y1:T ) from p (xt+1|y1:T )
Complete information about xt is given by these densities
An algorithm that can track these densities exactly is called anoptimal algorithm
In many practical situations, optimal algorithms are impossible toimplement
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 18 / 70
![Page 75: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/75.jpg)
Motivation for use of particle filtering
Example: Bearings-only tracking
Consider the tracking of an object based on bearings-only measurements
xt = Gxxt−1 + Guut
where xt = [x1,t x2,t x1,t x2,t ]>
Bearings-only range measurements are obtained by J sensors placed atknown locations in the sensor field
yj ,t = arctan
(x2,t − l2,jx1,t − l1,j
)+ vt(n)
Given the measurements of J sensors, yt = [y1,t y2,t · · · yJ,t ]>, andthe movement model of the object, the goal is to track the object intime, that is, estimate its position and velocity
In the literature this problem is known as target tracking
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 19 / 70
![Page 76: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/76.jpg)
Motivation for use of particle filtering
Example: Bearings-only tracking
Consider the tracking of an object based on bearings-only measurements
xt = Gxxt−1 + Guut
where xt = [x1,t x2,t x1,t x2,t ]>
Bearings-only range measurements are obtained by J sensors placed atknown locations in the sensor field
yj ,t = arctan
(x2,t − l2,jx1,t − l1,j
)+ vt(n)
Given the measurements of J sensors, yt = [y1,t y2,t · · · yJ,t ]>, andthe movement model of the object, the goal is to track the object intime, that is, estimate its position and velocity
In the literature this problem is known as target tracking
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 19 / 70
![Page 77: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/77.jpg)
Motivation for use of particle filtering
Example: Bearings-only tracking
Consider the tracking of an object based on bearings-only measurements
xt = Gxxt−1 + Guut
where xt = [x1,t x2,t x1,t x2,t ]>
Bearings-only range measurements are obtained by J sensors placed atknown locations in the sensor field
yj ,t = arctan
(x2,t − l2,jx1,t − l1,j
)+ vt(n)
Given the measurements of J sensors, yt = [y1,t y2,t · · · yJ,t ]>, andthe movement model of the object, the goal is to track the object intime, that is, estimate its position and velocity
In the literature this problem is known as target tracking
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 19 / 70
![Page 78: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/78.jpg)
Motivation for use of particle filtering
A target in a two-dimensional space and two BOT sensors
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 20 / 70
![Page 79: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/79.jpg)
Motivation for use of particle filtering
Recursion for filtering
Suppose that at time t − 1, we know the observations y1:t−1 and thea posteriori PDF p
(xt−1|y1:t−1
)Once yt becomes available, we update the PDF
p(xt |y1:t
)∝ p
(yt |xt
)p(xt |y1:t−1
)
The predictive density can be obtained as
p(xt |y1:t−1
)=
∫p(xt |xt−1
)p(xt−1|y1:t−1
)dxt−1
The recursive equation for updating the filtering density becomes
p(xt |y1:t
)∝ p
(yt |xt
)×
∫p(xt |xt−1
)p(xt−1|y1:t−1
)dxt−1
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 21 / 70
![Page 80: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/80.jpg)
Motivation for use of particle filtering
Recursion for filtering
Suppose that at time t − 1, we know the observations y1:t−1 and thea posteriori PDF p
(xt−1|y1:t−1
)Once yt becomes available, we update the PDF
p(xt |y1:t
)∝ p
(yt |xt
)p(xt |y1:t−1
)The predictive density can be obtained as
p(xt |y1:t−1
)=
∫p(xt |xt−1
)p(xt−1|y1:t−1
)dxt−1
The recursive equation for updating the filtering density becomes
p(xt |y1:t
)∝ p
(yt |xt
)×
∫p(xt |xt−1
)p(xt−1|y1:t−1
)dxt−1
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 21 / 70
![Page 81: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/81.jpg)
Motivation for use of particle filtering
Recursion for filtering
Suppose that at time t − 1, we know the observations y1:t−1 and thea posteriori PDF p
(xt−1|y1:t−1
)Once yt becomes available, we update the PDF
p(xt |y1:t
)∝ p
(yt |xt
)p(xt |y1:t−1
)The predictive density can be obtained as
p(xt |y1:t−1
)=
∫p(xt |xt−1
)p(xt−1|y1:t−1
)dxt−1
The recursive equation for updating the filtering density becomes
p(xt |y1:t
)∝ p
(yt |xt
)×
∫p(xt |xt−1
)p(xt−1|y1:t−1
)dxt−1
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 21 / 70
![Page 82: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/82.jpg)
Motivation for use of particle filtering
Recursion for filtering
Suppose that at time t − 1, we know the observations y1:t−1 and thea posteriori PDF p
(xt−1|y1:t−1
)Once yt becomes available, we update the PDF
p(xt |y1:t
)∝ p
(yt |xt
)p(xt |y1:t−1
)The predictive density can be obtained as
p(xt |y1:t−1
)=
∫p(xt |xt−1
)p(xt−1|y1:t−1
)dxt−1
The recursive equation for updating the filtering density becomes
p(xt |y1:t
)∝ p
(yt |xt
)×
∫p(xt |xt−1
)p(xt−1|y1:t−1
)dxt−1
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 21 / 70
![Page 83: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/83.jpg)
Motivation for use of particle filtering
Carrying out the filtering recursion
There are at least two problems in carrying out the above recursion
1 Solving the integral in order to obtain the predictive density
2 Combining the likelihood and the predictive density in order to getthe updated filtering density
In many problems the recursive evaluation of the densities of the statespace model cannot be done analytically, which entails that we haveto resort to numerical methods
PF can be employed with elegance and with performancecharacterized by high accuracy
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 22 / 70
![Page 84: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/84.jpg)
Motivation for use of particle filtering
Carrying out the filtering recursion
There are at least two problems in carrying out the above recursion
1 Solving the integral in order to obtain the predictive density
2 Combining the likelihood and the predictive density in order to getthe updated filtering density
In many problems the recursive evaluation of the densities of the statespace model cannot be done analytically, which entails that we haveto resort to numerical methods
PF can be employed with elegance and with performancecharacterized by high accuracy
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 22 / 70
![Page 85: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/85.jpg)
Motivation for use of particle filtering
Carrying out the filtering recursion
There are at least two problems in carrying out the above recursion
1 Solving the integral in order to obtain the predictive density
2 Combining the likelihood and the predictive density in order to getthe updated filtering density
In many problems the recursive evaluation of the densities of the statespace model cannot be done analytically, which entails that we haveto resort to numerical methods
PF can be employed with elegance and with performancecharacterized by high accuracy
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 22 / 70
![Page 86: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/86.jpg)
Motivation for use of particle filtering
Carrying out the filtering recursion
There are at least two problems in carrying out the above recursion
1 Solving the integral in order to obtain the predictive density
2 Combining the likelihood and the predictive density in order to getthe updated filtering density
In many problems the recursive evaluation of the densities of the statespace model cannot be done analytically, which entails that we haveto resort to numerical methods
PF can be employed with elegance and with performancecharacterized by high accuracy
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 22 / 70
![Page 87: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/87.jpg)
Motivation for use of particle filtering
Carrying out the filtering recursion
There are at least two problems in carrying out the above recursion
1 Solving the integral in order to obtain the predictive density
2 Combining the likelihood and the predictive density in order to getthe updated filtering density
In many problems the recursive evaluation of the densities of the statespace model cannot be done analytically, which entails that we haveto resort to numerical methods
PF can be employed with elegance and with performancecharacterized by high accuracy
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 22 / 70
![Page 88: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/88.jpg)
The basic idea
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 23 / 70
![Page 89: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/89.jpg)
The basic idea
Approximation of densities by discrete random measures
Densities can be approximated by discrete random measures
χ ={
x(m)t ,w
(m)t
}M
m=1
The discrete random measure approximates the density by
p(xt |y1:t) ≈M∑
m=1
w(m)t δ
(xt − x
(m)t
)
Computations of expectations simplify to summations
E (h(Xt)) =
∫h(xt)p(xt |y1:t)dxt
⇓
E (h(Xt)) ≈M∑
m=1
w(m)t h
(x
(m)t
)
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 24 / 70
![Page 90: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/90.jpg)
The basic idea
Approximation of densities by discrete random measures
Densities can be approximated by discrete random measures
χ ={
x(m)t ,w
(m)t
}M
m=1
The discrete random measure approximates the density by
p(xt |y1:t) ≈M∑
m=1
w(m)t δ
(xt − x
(m)t
)Computations of expectations simplify to summations
E (h(Xt)) =
∫h(xt)p(xt |y1:t)dxt
⇓
E (h(Xt)) ≈M∑
m=1
w(m)t h
(x
(m)t
)The Particle Filtering Methodology in Signal Processing Petar M. Djuric 24 / 70
![Page 91: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/91.jpg)
The basic idea
Approximation of densities by discrete random measures
Densities can be approximated by discrete random measures
χ ={
x(m)t ,w
(m)t
}M
m=1
The discrete random measure approximates the density by
p(xt |y1:t) ≈M∑
m=1
w(m)t δ
(xt − x
(m)t
)Computations of expectations simplify to summations
E (h(Xt)) =
∫h(xt)p(xt |y1:t)dxt
⇓
E (h(Xt)) ≈M∑
m=1
w(m)t h
(x
(m)t
)The Particle Filtering Methodology in Signal Processing Petar M. Djuric 24 / 70
![Page 92: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/92.jpg)
The basic idea
Example of approximation of densities
Assume that independent particles, x(m)t , can be drawn from p(xt |y1:t)
All the particles have the same weight
E (h(Xt)) =
∫h(xt)p(xt |y1:t)dxt → E (h(Xt)) =
1
M
M∑m=1
h(
x(m)t
)
E (·) is an unbiased estimator of the conditional expectation E (·)
If the variance σ2h <∞ → the variance of E (·) is σ2
E(h(·))=
σ2h
M
As M −→∞
Strong law of large numbers Central limit theorem
E (h(Xt))a.s.−→ E (h(Xt)) E (h(Xt))
d−→ N(
E (h(Xt)),σ2
hM
)
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 25 / 70
![Page 93: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/93.jpg)
The basic idea
Example of approximation of densities
Assume that independent particles, x(m)t , can be drawn from p(xt |y1:t)
All the particles have the same weight
E (h(Xt)) =
∫h(xt)p(xt |y1:t)dxt → E (h(Xt)) =
1
M
M∑m=1
h(
x(m)t
)E (·) is an unbiased estimator of the conditional expectation E (·)
If the variance σ2h <∞ → the variance of E (·) is σ2
E(h(·))=
σ2h
M
As M −→∞
Strong law of large numbers Central limit theorem
E (h(Xt))a.s.−→ E (h(Xt)) E (h(Xt))
d−→ N(
E (h(Xt)),σ2
hM
)
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 25 / 70
![Page 94: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/94.jpg)
The basic idea
Example of approximation of densities
Assume that independent particles, x(m)t , can be drawn from p(xt |y1:t)
All the particles have the same weight
E (h(Xt)) =
∫h(xt)p(xt |y1:t)dxt → E (h(Xt)) =
1
M
M∑m=1
h(
x(m)t
)E (·) is an unbiased estimator of the conditional expectation E (·)
If the variance σ2h <∞ → the variance of E (·) is σ2
E(h(·))=
σ2h
M
As M −→∞
Strong law of large numbers Central limit theorem
E (h(Xt))a.s.−→ E (h(Xt)) E (h(Xt))
d−→ N(
E (h(Xt)),σ2
hM
)
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 25 / 70
![Page 95: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/95.jpg)
The basic idea
Example of approximation of densities
Assume that independent particles, x(m)t , can be drawn from p(xt |y1:t)
All the particles have the same weight
E (h(Xt)) =
∫h(xt)p(xt |y1:t)dxt → E (h(Xt)) =
1
M
M∑m=1
h(
x(m)t
)E (·) is an unbiased estimator of the conditional expectation E (·)
If the variance σ2h <∞ → the variance of E (·) is σ2
E(h(·))=
σ2h
M
As M −→∞
Strong law of large numbers Central limit theorem
E (h(Xt))a.s.−→ E (h(Xt)) E (h(Xt))
d−→ N(
E (h(Xt)),σ2
hM
)The Particle Filtering Methodology in Signal Processing Petar M. Djuric 25 / 70
![Page 96: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/96.jpg)
The basic idea
Example of approximation of densities
Assume that independent particles, x(m)t , can be drawn from p(xt |y1:t)
All the particles have the same weight
E (h(Xt)) =
∫h(xt)p(xt |y1:t)dxt → E (h(Xt)) =
1
M
M∑m=1
h(
x(m)t
)E (·) is an unbiased estimator of the conditional expectation E (·)
If the variance σ2h <∞ → the variance of E (·) is σ2
E(h(·))=
σ2h
M
As M −→∞
Strong law of large numbers Central limit theorem
E (h(Xt))a.s.−→ E (h(Xt)) E (h(Xt))
d−→ N(
E (h(Xt)),σ2
hM
)The Particle Filtering Methodology in Signal Processing Petar M. Djuric 25 / 70
![Page 97: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/97.jpg)
The basic idea
The concept of importance sampling (IS)
In practice, we often cannot draw samples directly from p(xt |y1:t)
Particles are drawn from π(xt) and the estimate of E (h(Xt)) becomes
E (h(Xt)) ≈ 1
M
M∑m=1
w∗(m)t h
(x
(m)t
)or
E (h(Xt)) ≈M∑
m=1
w(m)t h
(x
(m)t
)
where w∗(m)t =
p(x(m)t |y1:t)
π(x(m)t )
and w(m)t = w
∗(m)t∑M
i=1 w∗(i)t
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 26 / 70
![Page 98: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/98.jpg)
The basic idea
The concept of importance sampling (IS)
In practice, we often cannot draw samples directly from p(xt |y1:t)
Particles are drawn from π(xt) and the estimate of E (h(Xt)) becomes
E (h(Xt)) ≈ 1
M
M∑m=1
w∗(m)t h
(x
(m)t
)or
E (h(Xt)) ≈M∑
m=1
w(m)t h
(x
(m)t
)
where w∗(m)t =
p(x(m)t |y1:t)
π(x(m)t )
and w(m)t = w
∗(m)t∑M
i=1 w∗(i)t
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 26 / 70
![Page 99: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/99.jpg)
The basic idea
Summary
We use a random measure to approximate p(xt |y1:t) in two steps:
1 Drawing samples from a proposal function π(xt), which needs to beknown only up to a multiplicative constant, i.e.,
x(m)t ∼ π(xt), m = 1, 2, · · · ,M
2 Computing the weights of the particles w(m)t
Particles
True posterior distribution
Proposal distributionWeights
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 27 / 70
![Page 100: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/100.jpg)
The basic idea
How do we obtain χt from χt−1?
We assume χt−1 ={
x(m)t−1,w
(m)t−1
}M
m=1is given.
Step one: Generation of new particles from a proposal function, i.e.,
x(m)t ∼ π (xt) , m = 1, · · · ,M
and augmentation of the particle stream x(m)0:t−1 with x
(m)t
Step two: Computation of the particle weights of x(m)t ,
m = 1, · · · ,M, i.e.
w(m)t ∝ w
(m)t−1 × update factor, m = 1, · · · ,M
and normalization of the weights
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 28 / 70
![Page 101: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/101.jpg)
The basic idea
How do we obtain χt from χt−1?
We assume χt−1 ={
x(m)t−1,w
(m)t−1
}M
m=1is given.
Step one: Generation of new particles from a proposal function, i.e.,
x(m)t ∼ π (xt) , m = 1, · · · ,M
and augmentation of the particle stream x(m)0:t−1 with x
(m)t
Step two: Computation of the particle weights of x(m)t ,
m = 1, · · · ,M, i.e.
w(m)t ∝ w
(m)t−1 × update factor, m = 1, · · · ,M
and normalization of the weights
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 28 / 70
![Page 102: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/102.jpg)
The basic idea
Mathematical formulation of the algorithm
Particle generationBasis
π (x0:t |y1:t) = π (xt |x0:t−1, y1:t)π(x0:t−1|y1:t−1
)x
(m)0:t−1 ∼ π
(x0:t−1|y1:t−1
)w
(m)t−1 ∝
p(x(m)0:t−1|yt−1)
π(x(m)0:t−1|y1:t−1)
Augmentation of the trajectory x(m)0:t−1 with x
(m)t
x(m)t ∼ π
(xt |x(m)
0:t−1, y1:t
), m = 1, · · · ,M
Weight update
w(m)t ∝
p(yt |x
(m)t
)p(x(m)
t |x(m)t−1
)π(x(m)
t |x(m)0:t−1,y1:t
) w(m)t−1, m = 1, · · · ,M
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 29 / 70
![Page 103: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/103.jpg)
The basic idea
Example: random walk
Consider the state-space model
xt = xt−1 + ut
yt = xt + vt
0 10 20 30 40 50 60 70 80 90 −12
−10
−8
−6
−4
−2
0
2
time n
stat
e x(
n)
0 10 20 30 40 50 60 70 80 90 −12
−10
−8
−6
−4
−2
0
2
time n
obse
rvat
ion
y(n)
t = 0 t = 1 t = 2xt 1.44 0.89 −1.28yt 0.96 0.98 −0.51
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 30 / 70
![Page 104: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/104.jpg)
The basic idea
Example: SIR for random walk
We choose as a proposal function
π(xt) = p(xt |xt−1)
Therefore, we generate the particles according to
x(m)t ∼ p(xt |x (m)
t−1), m = 1, 2, · · · ,M
The update of the weights is done according to
w(m)t ∝ p(yt |x (m)
t )w(m)t−1
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 31 / 70
![Page 105: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/105.jpg)
The basic idea
Example: SIR for random walk
We choose as a proposal function
π(xt) = p(xt |xt−1)
Therefore, we generate the particles according to
x(m)t ∼ p(xt |x (m)
t−1), m = 1, 2, · · · ,M
The update of the weights is done according to
w(m)t ∝ p(yt |x (m)
t )w(m)t−1
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 31 / 70
![Page 106: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/106.jpg)
The basic idea
Example: SIR for random walk
We choose as a proposal function
π(xt) = p(xt |xt−1)
Therefore, we generate the particles according to
x(m)t ∼ p(xt |x (m)
t−1), m = 1, 2, · · · ,M
The update of the weights is done according to
w(m)t ∝ p(yt |x (m)
t )w(m)t−1
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 31 / 70
![Page 107: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/107.jpg)
The basic idea
Example: Step by step SIR for random walk
Initialization: t = 0
x(m)0 ∼ N (0, 1) m = 1, 2, . . . ,M.
Note that all the weights are equal
−3 −2 −1 0 1 2 30
0.001
0.002
0.003
0.004
0.005
0.006
0.007
0.008
0.009
0.01
particles x(m)(0)
wei
ghts
w(m
) (0)
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 32 / 70
![Page 108: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/108.jpg)
The basic idea
Example: Step by step SIR for random walk (cont.)
Time instant: t = 1
Generation of particles using the proposal function:
x(m)1 ∼ p
(x1|x (m)
0
)=
1√2π
exp
(−
(x1 − x(m)0 )2
2
)Weight update:
w(m)1 ∝ exp
(−
(y1 − x(m)1 )2
2
)
−5 0 50
0.002
0.004
0.006
0.008
0.01
0.012
0.014
particles x(m)(1)
wei
ghts
w(m
) (1)
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 33 / 70
![Page 109: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/109.jpg)
The basic idea
Example: Step by step SIR for random walk (cont.)
Time instant: t = 2
Generation of particles using the proposal function:
x(m)2 ∼ p
(x2|x (m)
1
)=
1√2π
exp
(−
(x2 − x(m)1 )2
2
)Weight update:
w(m)2 ∝ w
(m)1 exp
(−
(y2 − x(m)2 )2
2
)
−6 −4 −2 0 2 4 60
0.005
0.01
0.015
0.02
0.025
particles x(m)(2)
wei
ghts
w(m
) (2)
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 34 / 70
![Page 110: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/110.jpg)
The basic idea
Choice of proposal function
The proposal (importance) function plays a crucial role in theperformance of PF
It is desirable to use easy-to-sample proposal functions that produceparticles with a large enough variance in order to avoid exploration ofthe state space in too narrow regions and thereby contributing tolosing the tracks of the state, but not too large to alleviate generationof too dispersed particles
The support of the proposal functions has to be the same as that ofthe targeted distribution
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 35 / 70
![Page 111: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/111.jpg)
The basic idea
Choice of proposal function
The proposal (importance) function plays a crucial role in theperformance of PF
It is desirable to use easy-to-sample proposal functions that produceparticles with a large enough variance in order to avoid exploration ofthe state space in too narrow regions and thereby contributing tolosing the tracks of the state, but not too large to alleviate generationof too dispersed particles
The support of the proposal functions has to be the same as that ofthe targeted distribution
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 35 / 70
![Page 112: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/112.jpg)
The basic idea
Choice of proposal function
The proposal (importance) function plays a crucial role in theperformance of PF
It is desirable to use easy-to-sample proposal functions that produceparticles with a large enough variance in order to avoid exploration ofthe state space in too narrow regions and thereby contributing tolosing the tracks of the state, but not too large to alleviate generationof too dispersed particles
The support of the proposal functions has to be the same as that ofthe targeted distribution
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 35 / 70
![Page 113: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/113.jpg)
The basic idea
The optimal importance function
The optimal choice for the importance function is the true posteriordistribution, i.e.,
π(
xt |x(m)0:t−1, y1:t
)= p
(xt |x(m)
t−1, yt
)This choice corresponds to the weight calculation
w(m)t ∝ w
(m)t−1p
(yt |x
(m)t−1
)Advantage: this importance function minimizes the variance of theweights
Disadvantage: it is difficult to sample from p(
xt |x(m)t−1, yt
)and the
update of the weights requires the computation of the integral
p(
yt | x(m)t−1
)=
∫p(yt | xt)p
(xt | x(m)
t−1
)dxt
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 36 / 70
![Page 114: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/114.jpg)
The basic idea
The optimal importance function
The optimal choice for the importance function is the true posteriordistribution, i.e.,
π(
xt |x(m)0:t−1, y1:t
)= p
(xt |x(m)
t−1, yt
)This choice corresponds to the weight calculation
w(m)t ∝ w
(m)t−1p
(yt |x
(m)t−1
)Advantage: this importance function minimizes the variance of theweights
Disadvantage: it is difficult to sample from p(
xt |x(m)t−1, yt
)and the
update of the weights requires the computation of the integral
p(
yt | x(m)t−1
)=
∫p(yt | xt)p
(xt | x(m)
t−1
)dxt
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 36 / 70
![Page 115: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/115.jpg)
The basic idea
The optimal importance function
The optimal choice for the importance function is the true posteriordistribution, i.e.,
π(
xt |x(m)0:t−1, y1:t
)= p
(xt |x(m)
t−1, yt
)This choice corresponds to the weight calculation
w(m)t ∝ w
(m)t−1p
(yt |x
(m)t−1
)Advantage: this importance function minimizes the variance of theweights
Disadvantage: it is difficult to sample from p(
xt |x(m)t−1, yt
)and the
update of the weights requires the computation of the integral
p(
yt | x(m)t−1
)=
∫p(yt | xt)p
(xt | x(m)
t−1
)dxt
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 36 / 70
![Page 116: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/116.jpg)
The basic idea
The prior importance function
A popular choice for the importance function is the prior
π(
xt |x(m)0:t−1, y1:t
)= p
(xt |x(m)
t−1
)This choice corresponds to the weight calculation
w(m)t ∝ w
(m)t−1 p
(yt |x
(m)t
)
Advantage: the computation of the weights is easy
Disadvantage: the generation of the particles is implemented withoutthe use of observations
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 37 / 70
![Page 117: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/117.jpg)
The basic idea
The prior importance function
A popular choice for the importance function is the prior
π(
xt |x(m)0:t−1, y1:t
)= p
(xt |x(m)
t−1
)This choice corresponds to the weight calculation
w(m)t ∝ w
(m)t−1 p
(yt |x
(m)t
)
Advantage: the computation of the weights is easy
Disadvantage: the generation of the particles is implemented withoutthe use of observations
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 37 / 70
![Page 118: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/118.jpg)
The basic idea
The prior importance function
A popular choice for the importance function is the prior
π(
xt |x(m)0:t−1, y1:t
)= p
(xt |x(m)
t−1
)This choice corresponds to the weight calculation
w(m)t ∝ w
(m)t−1 p
(yt |x
(m)t
)
Advantage: the computation of the weights is easy
Disadvantage: the generation of the particles is implemented withoutthe use of observations
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 37 / 70
![Page 119: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/119.jpg)
The basic idea
Degeneracy of the random measure: Resampling
In particle filtering the discrete random measure degenerates quickly andonly few particles are assigned meaningful weights
��������������������������������������������������
��������������������������������������������������time instant
random measure
�x(m)t ; w(m)t Mm=1tt+ 1t+ 2 �x(m)t+1 ; w(m)t+1Mm=1
�x(m)t+2 ; w(m)t+2Mm=1
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 38 / 70
![Page 120: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/120.jpg)
The basic idea
Resampling
A measure of this degeneracy is the effective particle size
Meff =M
1 + Var(
w∗(m)t
) → Meff =1∑M
m=1
(w
(m)t
)2
Degeneracy is reduced by using good importance sampling functionsand resampling
Resampling eliminates particles with small weights and replicates theones with large weights
1/5
1
0
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 39 / 70
![Page 121: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/121.jpg)
The basic idea
Pictorial description of resampling
�����������������������������������������
�����������������������������������������
random measure
time instant
particles after resamplingand propagation
particles after resamplingand propagationt+ 1
t particles from step t� 1�x(m)t ; w(m)t Mm=1�x(m)t+1 ; w(m)t+1Mm=1
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 40 / 70
![Page 122: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/122.jpg)
The basic idea
Flowchart of the SIR algorithm
of desired unknowns
new observation
yes
Exit
Update weights
no
observations?More
Normalize weights
M21 ...
... M2
Propose particles
Initializeweights
Output
Compute estimates
isno
yes
evaluate effectiveparticle size
resampling
necessary?resampling
1
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 41 / 70
![Page 123: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/123.jpg)
Some particle filtering methods
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 42 / 70
![Page 124: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/124.jpg)
Some particle filtering methods
SIR PF
The SIR method is the simplest of all the particle filtering methods
It was named bootstrap filter
It employs the prior density for drawing particles, which implies thatthe weights are only proportional to the likelihood of the drawnparticles
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 43 / 70
![Page 125: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/125.jpg)
Some particle filtering methods
SIR PF
The SIR method is the simplest of all the particle filtering methods
It was named bootstrap filter
It employs the prior density for drawing particles, which implies thatthe weights are only proportional to the likelihood of the drawnparticles
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 43 / 70
![Page 126: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/126.jpg)
Some particle filtering methods
SIR PF
The SIR method is the simplest of all the particle filtering methods
It was named bootstrap filter
It employs the prior density for drawing particles, which implies thatthe weights are only proportional to the likelihood of the drawnparticles
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 43 / 70
![Page 127: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/127.jpg)
Some particle filtering methods
Auxiliary particle filtering (APF)
It explores the state space by using the latest measurements.
Particles are propagated using the prior
Weights of the newly generated particles are evaluated by the latestmeasurement
Resampling is applied using the so computed weights
The resampled streams are propagated by generating new particles
The weights of the new particles are computed
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 44 / 70
![Page 128: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/128.jpg)
Some particle filtering methods
Auxiliary particle filtering (APF)
It explores the state space by using the latest measurements.
Particles are propagated using the prior
Weights of the newly generated particles are evaluated by the latestmeasurement
Resampling is applied using the so computed weights
The resampled streams are propagated by generating new particles
The weights of the new particles are computed
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 44 / 70
![Page 129: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/129.jpg)
Some particle filtering methods
Auxiliary particle filtering (APF)
It explores the state space by using the latest measurements.
Particles are propagated using the prior
Weights of the newly generated particles are evaluated by the latestmeasurement
Resampling is applied using the so computed weights
The resampled streams are propagated by generating new particles
The weights of the new particles are computed
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 44 / 70
![Page 130: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/130.jpg)
Some particle filtering methods
Auxiliary particle filtering (APF)
It explores the state space by using the latest measurements.
Particles are propagated using the prior
Weights of the newly generated particles are evaluated by the latestmeasurement
Resampling is applied using the so computed weights
The resampled streams are propagated by generating new particles
The weights of the new particles are computed
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 44 / 70
![Page 131: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/131.jpg)
Some particle filtering methods
Auxiliary particle filtering (APF)
It explores the state space by using the latest measurements.
Particles are propagated using the prior
Weights of the newly generated particles are evaluated by the latestmeasurement
Resampling is applied using the so computed weights
The resampled streams are propagated by generating new particles
The weights of the new particles are computed
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 44 / 70
![Page 132: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/132.jpg)
Some particle filtering methods
Unscented particle filtering (UPF)
The method is based on the unscented transform
Rather than linearizing the model, the UKF uses a small deterministicgrid of points (called sigma points) which when propagated throughthe system capture accurately both the mean and the covariance
In the context of particle filtering, the UKF computes Gaussianswhich are then used as importance functions
Each particle stream has its own UKF
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 45 / 70
![Page 133: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/133.jpg)
Some particle filtering methods
Unscented particle filtering (UPF)
The method is based on the unscented transform
Rather than linearizing the model, the UKF uses a small deterministicgrid of points (called sigma points) which when propagated throughthe system capture accurately both the mean and the covariance
In the context of particle filtering, the UKF computes Gaussianswhich are then used as importance functions
Each particle stream has its own UKF
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 45 / 70
![Page 134: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/134.jpg)
Some particle filtering methods
Unscented particle filtering (UPF)
The method is based on the unscented transform
Rather than linearizing the model, the UKF uses a small deterministicgrid of points (called sigma points) which when propagated throughthe system capture accurately both the mean and the covariance
In the context of particle filtering, the UKF computes Gaussianswhich are then used as importance functions
Each particle stream has its own UKF
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 45 / 70
![Page 135: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/135.jpg)
Some particle filtering methods
Unscented particle filtering (UPF)
The method is based on the unscented transform
Rather than linearizing the model, the UKF uses a small deterministicgrid of points (called sigma points) which when propagated throughthe system capture accurately both the mean and the covariance
In the context of particle filtering, the UKF computes Gaussianswhich are then used as importance functions
Each particle stream has its own UKF
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 45 / 70
![Page 136: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/136.jpg)
Some particle filtering methods
Gaussian particle filtering (GPF)
The posterior and the predictive densities are approximated withGaussians as is done by the extended Kalman filter (EKF)
However, unlike with the EKF, we do not linearize any of thenonlinearities in the system
Its advantage over other PF methods is that it does not requireresampling
Another advantage is that the estimation of constant parameters isnot a problem as is the case with other PF methods
Its disadvantage is that if the approximated densities are notGaussians, the estimates may be inaccurate and the filter may diverge
An alternative could be the Gaussian sum particle filtering where thedensities in the system are approximated by mixture Gaussians
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 46 / 70
![Page 137: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/137.jpg)
Some particle filtering methods
Gaussian particle filtering (GPF)
The posterior and the predictive densities are approximated withGaussians as is done by the extended Kalman filter (EKF)
However, unlike with the EKF, we do not linearize any of thenonlinearities in the system
Its advantage over other PF methods is that it does not requireresampling
Another advantage is that the estimation of constant parameters isnot a problem as is the case with other PF methods
Its disadvantage is that if the approximated densities are notGaussians, the estimates may be inaccurate and the filter may diverge
An alternative could be the Gaussian sum particle filtering where thedensities in the system are approximated by mixture Gaussians
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 46 / 70
![Page 138: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/138.jpg)
Some particle filtering methods
Gaussian particle filtering (GPF)
The posterior and the predictive densities are approximated withGaussians as is done by the extended Kalman filter (EKF)
However, unlike with the EKF, we do not linearize any of thenonlinearities in the system
Its advantage over other PF methods is that it does not requireresampling
Another advantage is that the estimation of constant parameters isnot a problem as is the case with other PF methods
Its disadvantage is that if the approximated densities are notGaussians, the estimates may be inaccurate and the filter may diverge
An alternative could be the Gaussian sum particle filtering where thedensities in the system are approximated by mixture Gaussians
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 46 / 70
![Page 139: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/139.jpg)
Some particle filtering methods
Gaussian particle filtering (GPF)
The posterior and the predictive densities are approximated withGaussians as is done by the extended Kalman filter (EKF)
However, unlike with the EKF, we do not linearize any of thenonlinearities in the system
Its advantage over other PF methods is that it does not requireresampling
Another advantage is that the estimation of constant parameters isnot a problem as is the case with other PF methods
Its disadvantage is that if the approximated densities are notGaussians, the estimates may be inaccurate and the filter may diverge
An alternative could be the Gaussian sum particle filtering where thedensities in the system are approximated by mixture Gaussians
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 46 / 70
![Page 140: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/140.jpg)
Some particle filtering methods
Gaussian particle filtering (GPF)
The posterior and the predictive densities are approximated withGaussians as is done by the extended Kalman filter (EKF)
However, unlike with the EKF, we do not linearize any of thenonlinearities in the system
Its advantage over other PF methods is that it does not requireresampling
Another advantage is that the estimation of constant parameters isnot a problem as is the case with other PF methods
Its disadvantage is that if the approximated densities are notGaussians, the estimates may be inaccurate and the filter may diverge
An alternative could be the Gaussian sum particle filtering where thedensities in the system are approximated by mixture Gaussians
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 46 / 70
![Page 141: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/141.jpg)
Some particle filtering methods
Gaussian particle filtering (GPF)
The posterior and the predictive densities are approximated withGaussians as is done by the extended Kalman filter (EKF)
However, unlike with the EKF, we do not linearize any of thenonlinearities in the system
Its advantage over other PF methods is that it does not requireresampling
Another advantage is that the estimation of constant parameters isnot a problem as is the case with other PF methods
Its disadvantage is that if the approximated densities are notGaussians, the estimates may be inaccurate and the filter may diverge
An alternative could be the Gaussian sum particle filtering where thedensities in the system are approximated by mixture Gaussians
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 46 / 70
![Page 142: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/142.jpg)
Some particle filtering methods
Comparison of methods – Bearings only tracking
0 50 100 150 200 250 30010
−1
100
101
102
103
t
MS
E P
ositi
on
SPFAUXGPFUPF
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 47 / 70
![Page 143: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/143.jpg)
Handling constant parameters
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 48 / 70
![Page 144: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/144.jpg)
Handling constant parameters
Handling constant parameters
The PF methodology was originally devised for the estimation ofdynamic signals rather than static parameters
The most efficient way to address the problem is to integrate out theunknown parameters when possible, either analytically or by MonteCarlo procedures
A common feature of these approaches is that they introduce artificialevolution of the fixed parameters
A recent work introduced a special class of PF called density assistedparticle filters that approximate the filtering density with a predefinedparametric density by generalizing the concepts of Gaussian particlefilters and Gaussian sum particle filters
These new filters can cope with constant parameters more naturallythan previously proposed methods
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 49 / 70
![Page 145: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/145.jpg)
Handling constant parameters
Handling constant parameters
The PF methodology was originally devised for the estimation ofdynamic signals rather than static parameters
The most efficient way to address the problem is to integrate out theunknown parameters when possible, either analytically or by MonteCarlo procedures
A common feature of these approaches is that they introduce artificialevolution of the fixed parameters
A recent work introduced a special class of PF called density assistedparticle filters that approximate the filtering density with a predefinedparametric density by generalizing the concepts of Gaussian particlefilters and Gaussian sum particle filters
These new filters can cope with constant parameters more naturallythan previously proposed methods
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 49 / 70
![Page 146: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/146.jpg)
Handling constant parameters
Handling constant parameters
The PF methodology was originally devised for the estimation ofdynamic signals rather than static parameters
The most efficient way to address the problem is to integrate out theunknown parameters when possible, either analytically or by MonteCarlo procedures
A common feature of these approaches is that they introduce artificialevolution of the fixed parameters
A recent work introduced a special class of PF called density assistedparticle filters that approximate the filtering density with a predefinedparametric density by generalizing the concepts of Gaussian particlefilters and Gaussian sum particle filters
These new filters can cope with constant parameters more naturallythan previously proposed methods
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 49 / 70
![Page 147: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/147.jpg)
Handling constant parameters
Handling constant parameters
The PF methodology was originally devised for the estimation ofdynamic signals rather than static parameters
The most efficient way to address the problem is to integrate out theunknown parameters when possible, either analytically or by MonteCarlo procedures
A common feature of these approaches is that they introduce artificialevolution of the fixed parameters
A recent work introduced a special class of PF called density assistedparticle filters that approximate the filtering density with a predefinedparametric density by generalizing the concepts of Gaussian particlefilters and Gaussian sum particle filters
These new filters can cope with constant parameters more naturallythan previously proposed methods
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 49 / 70
![Page 148: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/148.jpg)
Handling constant parameters
Handling constant parameters
The PF methodology was originally devised for the estimation ofdynamic signals rather than static parameters
The most efficient way to address the problem is to integrate out theunknown parameters when possible, either analytically or by MonteCarlo procedures
A common feature of these approaches is that they introduce artificialevolution of the fixed parameters
A recent work introduced a special class of PF called density assistedparticle filters that approximate the filtering density with a predefinedparametric density by generalizing the concepts of Gaussian particlefilters and Gaussian sum particle filters
These new filters can cope with constant parameters more naturallythan previously proposed methods
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 49 / 70
![Page 149: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/149.jpg)
Handling constant parameters
Kernel-based auxiliary particle filter
The inclusion of fixed parameters in the model implies extending therandom measure to the form
χt ={
x(m)t ,θ
(m)t ,w
(m)t
}M
m=1
The random measure approximates the density of interestp(xt ,θ | yt), which can be decomposed as
p(xt ,θ | y1:t) ∝ p(yt | xt ,θ)p(xt | θ, y1:t)p(θ | y1:t)
It is clear that there is a need for approximation of the density,p(θ | y1:t), e.g., by
p(θ | y1:t) ≈M∑
m=1
w(m)t N
(θ | θ(m)
t , h2Σθ,t
)
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 50 / 70
![Page 150: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/150.jpg)
Handling constant parameters
Kernel-based auxiliary particle filter
The inclusion of fixed parameters in the model implies extending therandom measure to the form
χt ={
x(m)t ,θ
(m)t ,w
(m)t
}M
m=1
The random measure approximates the density of interestp(xt ,θ | yt), which can be decomposed as
p(xt ,θ | y1:t) ∝ p(yt | xt ,θ)p(xt | θ, y1:t)p(θ | y1:t)
It is clear that there is a need for approximation of the density,p(θ | y1:t), e.g., by
p(θ | y1:t) ≈M∑
m=1
w(m)t N
(θ | θ(m)
t , h2Σθ,t
)
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 50 / 70
![Page 151: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/151.jpg)
Handling constant parameters
Kernel-based auxiliary particle filter
The inclusion of fixed parameters in the model implies extending therandom measure to the form
χt ={
x(m)t ,θ
(m)t ,w
(m)t
}M
m=1
The random measure approximates the density of interestp(xt ,θ | yt), which can be decomposed as
p(xt ,θ | y1:t) ∝ p(yt | xt ,θ)p(xt | θ, y1:t)p(θ | y1:t)
It is clear that there is a need for approximation of the density,p(θ | y1:t), e.g., by
p(θ | y1:t) ≈M∑
m=1
w(m)t N
(θ | θ(m)
t , h2Σθ,t
)
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 50 / 70
![Page 152: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/152.jpg)
Handling constant parameters
Density assisted particle filter
The approximating densities can be other than Gaussians or mixtures ofGaussians
1 Draw particles of θ from p(θ|y0:t−1), i.e.,
θ(m)t−1 ∼ p(θ|y0:t−1)
2 Draw particles according to
x(m)t ∼ p(xt |x(m)
t−1,θ(m)t−1)
3 Set θ(m)t = θ
(m)t−1
4 Update and normalize the weights
w(m)t ∝ p(yt |x
(m)t ,θ
(m)t )
5 Estimate the parameters of the density p(θ|y0:t) from
χt ={
x(m)t ,θ
(m)t ,w
(m)t
}M
m=1
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 51 / 70
![Page 153: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/153.jpg)
Handling constant parameters
Handling constant parameters – example
xt = axt−1 + ut
yt = bxt + vt
0 10 20 30 40 50 60 70 80 90 100−8
−6
−4
−2
0
2
4
t
x(t)
true stateDAPFLW
0 10 20 30 40 50 60 70 80 90 100−8
−6
−4
−2
0
2
4
6
t
x(t)
true stateDAPFLW
0 10 20 30 40 50 60 70 80 90 1000
0.5
1
t
MS
E(x
(t))
DAPFLW
0 10 20 30 40 50 60 70 80 90 1000
0.05
0.1
0.15
0.2
t
MS
E(a
)
DAPFLW
0 10 20 30 40 50 60 70 80 90 1000
0.2
0.4
0.6
0.8
t
MS
E(b
)
DAPFLW
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 52 / 70
![Page 154: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/154.jpg)
Rao-Blackwellization
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 53 / 70
![Page 155: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/155.jpg)
Rao-Blackwellization
An efficient approach: Rao-Blackwellized PF
Rao-Blackwellization allows for marginalization of part of the stateswhich are conditionally linear on the remaining states
This approach leads to a more accurate performance of SPF since asmaller state space is explored
Standard implementation of Rao-Blackwellization requires running asmany Kalman filters as particle streams
An efficient realization of the Rao-Blackwellized PF with only oneKalman filter has also been proposed
The new approach shows considerable computational savings withoutsacrificing tracking performance
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 54 / 70
![Page 156: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/156.jpg)
Rao-Blackwellization
An efficient approach: Rao-Blackwellized PF
Rao-Blackwellization allows for marginalization of part of the stateswhich are conditionally linear on the remaining states
This approach leads to a more accurate performance of SPF since asmaller state space is explored
Standard implementation of Rao-Blackwellization requires running asmany Kalman filters as particle streams
An efficient realization of the Rao-Blackwellized PF with only oneKalman filter has also been proposed
The new approach shows considerable computational savings withoutsacrificing tracking performance
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 54 / 70
![Page 157: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/157.jpg)
Rao-Blackwellization
An efficient approach: Rao-Blackwellized PF
Rao-Blackwellization allows for marginalization of part of the stateswhich are conditionally linear on the remaining states
This approach leads to a more accurate performance of SPF since asmaller state space is explored
Standard implementation of Rao-Blackwellization requires running asmany Kalman filters as particle streams
An efficient realization of the Rao-Blackwellized PF with only oneKalman filter has also been proposed
The new approach shows considerable computational savings withoutsacrificing tracking performance
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 54 / 70
![Page 158: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/158.jpg)
Rao-Blackwellization
An efficient approach: Rao-Blackwellized PF
Rao-Blackwellization allows for marginalization of part of the stateswhich are conditionally linear on the remaining states
This approach leads to a more accurate performance of SPF since asmaller state space is explored
Standard implementation of Rao-Blackwellization requires running asmany Kalman filters as particle streams
An efficient realization of the Rao-Blackwellized PF with only oneKalman filter has also been proposed
The new approach shows considerable computational savings withoutsacrificing tracking performance
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 54 / 70
![Page 159: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/159.jpg)
Rao-Blackwellization
An efficient approach: Rao-Blackwellized PF
Rao-Blackwellization allows for marginalization of part of the stateswhich are conditionally linear on the remaining states
This approach leads to a more accurate performance of SPF since asmaller state space is explored
Standard implementation of Rao-Blackwellization requires running asmany Kalman filters as particle streams
An efficient realization of the Rao-Blackwellized PF with only oneKalman filter has also been proposed
The new approach shows considerable computational savings withoutsacrificing tracking performance
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 54 / 70
![Page 160: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/160.jpg)
Rao-Blackwellization
Example: Bearings-only tracking with biased measurements
Consider K targets moving along a 2D sensor field of N sensors
xt = Gxxt−1 + Guut
The n−th sensor at time instant t measures the bearing information of thetargets sensed in the field
yn,t = gn(xt) + bn + vn,t n = 1, · · · ,N
gn(xt) =[arctan
(x2,1,t−x2,n
x1,1,t−x1,n
), · · · , arctan
(x2,K ,t−x2,n
x1,K ,t−x1,n
)]>where bn represents the unknown bias of the n−th sensor
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 55 / 70
![Page 161: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/161.jpg)
Rao-Blackwellization
Comparison of algorithms
Averaged MSE in m2 of two targets obtained by different methods usingM = 500 particles and J = 50 independent runs
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 56 / 70
![Page 162: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/162.jpg)
Prediction and smoothing
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 57 / 70
![Page 163: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/163.jpg)
Prediction and smoothing
Prediction
The prediction problem revolves around the estimation of the predictivedensity p(xt+l |y1:t), where l > 1
The approximation of the predictive density p(xt+l |y1:t) is obtained by
1 drawing particles x(m)t+1 from p(xt+1|x(m)
t )
2 drawing particles x(m)t+2 from p(xt+2|x(m)
t+1)
3 · · ·4 drawing particles x
(m)t+l from p(xt+l |x
(m)t+l−1)
5 and associating with the samples x(m)t+l the weights w
(m)t and thereby
forming the random measure {x(m)t+l ,w
(m)t }Mm=1.
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 58 / 70
![Page 164: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/164.jpg)
Prediction and smoothing
Smoothing
All the information about xt in this case is in the PDF p(xt |y1:T )
Forward PFRun PF in the forward direction and store all the random measures
χt = {x(m)t ,w
(m)t }Mm=1, n = 1, 2, · · · ,N
Backward recursionsSet the smoothing weights
w(m)s,T = w
(m)T and χs,T = {x(m)
T ,w(m)s,T }
Mm=1
For n = N − 1, · · · , 1, 0Computation of the smoothing weights ws,t
Compute the smoothing weights of x(m)t by
w(m)s,t =
∑Mj=1 w
(j)s,t+1
w(m)t p(x(j)
t+1|x(m)t )∑M
l=1 w(l)t p(x(j)
t+1|x(l)t )
Construction of the smoothing random measure χs,t
Set χs,t = {x(m)t ,w
(m)s,t }Mm=1
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 59 / 70
![Page 165: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/165.jpg)
Multiple particle filters
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 60 / 70
![Page 166: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/166.jpg)
Multiple particle filters
Multiple particle filters
We are interested in particle filtering methods for complex systems thatcan be represented by the following state-space model:
xt = fx(xt−1,ut)
yt = fy (xt , vt)
We assume that the dimension of xt is high
From practice we know that high-dimensional states would, in general,require a very large number of particles for accurate tracking of theposterior pdf of xt
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 61 / 70
![Page 167: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/167.jpg)
Multiple particle filters
Multiple particle filters
The underlying idea is to use multiple particle filters that communicateinformation about their posterior pdfs
The particle filters operate on partitioned subspaces of the complete statespace
The state space is decomposed into separate subspaces of lowerdimensionality which form a partition of the original space
We assume that the interest is in finding the marginal posterior densitiesof the state vectors that span these subspaces
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 62 / 70
![Page 168: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/168.jpg)
Multiple particle filters
Multiple particle filters
A system of multiple particle filters.
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 63 / 70
![Page 169: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/169.jpg)
Multiple particle filters
0 50 100 150 200 2500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
e (m)
F(e
)
SPFMPF 1MPF 2
CDFs of RMSEs (M = 1000, Mk = 500)
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 64 / 70
![Page 170: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/170.jpg)
Convergence issues
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 65 / 70
![Page 171: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/171.jpg)
Convergence issues
Convergence issues
PF is based on approximations of PDFs by discrete random measures
We expect that the approximations deteriorate as the dimension ofthe state space increases but hope that as the number of particlesM →∞, the approximation of the density will improve
The approximations are based on particle streams, and they aredependent because of the necessity for resampling
Therefore classical limit theorems for studying convergence cannot beapplied, which makes the analysis more difficult
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 66 / 70
![Page 172: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/172.jpg)
Convergence issues
Convergence issues
PF is based on approximations of PDFs by discrete random measures
We expect that the approximations deteriorate as the dimension ofthe state space increases but hope that as the number of particlesM →∞, the approximation of the density will improve
The approximations are based on particle streams, and they aredependent because of the necessity for resampling
Therefore classical limit theorems for studying convergence cannot beapplied, which makes the analysis more difficult
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 66 / 70
![Page 173: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/173.jpg)
Convergence issues
Convergence issues
PF is based on approximations of PDFs by discrete random measures
We expect that the approximations deteriorate as the dimension ofthe state space increases but hope that as the number of particlesM →∞, the approximation of the density will improve
The approximations are based on particle streams, and they aredependent because of the necessity for resampling
Therefore classical limit theorems for studying convergence cannot beapplied, which makes the analysis more difficult
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 66 / 70
![Page 174: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/174.jpg)
Convergence issues
Convergence issues
PF is based on approximations of PDFs by discrete random measures
We expect that the approximations deteriorate as the dimension ofthe state space increases but hope that as the number of particlesM →∞, the approximation of the density will improve
The approximations are based on particle streams, and they aredependent because of the necessity for resampling
Therefore classical limit theorems for studying convergence cannot beapplied, which makes the analysis more difficult
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 66 / 70
![Page 175: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/175.jpg)
Convergence issues
Convergence theorems
Theorem 1: If p(xt |xt−1) satisfies some mild conditions and thelikelihood function p(yt |xt) is bounded, continuous and strictlypositive, then
limM→∞
χMt = p(xt |y1:t)
almost surely
Theorem 2: If the likelihood function p(yt |xt) is bounded, for allt ≥ 1 there exists a constant ct independent of M such that for anybounded function
E (e2t ) ≤ ct
‖h(xt)‖2
M
where E (·) is an expectation operator, h(xt) is a function of xt , and‖h(xt)‖ denotes the supremum norm, i.e.,
‖h(xt‖ = supxt
|h(xt)|
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 67 / 70
![Page 176: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/176.jpg)
Convergence issues
Convergence theorems
Theorem 1: If p(xt |xt−1) satisfies some mild conditions and thelikelihood function p(yt |xt) is bounded, continuous and strictlypositive, then
limM→∞
χMt = p(xt |y1:t)
almost surely
Theorem 2: If the likelihood function p(yt |xt) is bounded, for allt ≥ 1 there exists a constant ct independent of M such that for anybounded function
E (e2t ) ≤ ct
‖h(xt)‖2
M
where E (·) is an expectation operator, h(xt) is a function of xt , and‖h(xt)‖ denotes the supremum norm, i.e.,
‖h(xt‖ = supxt
|h(xt)|
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 67 / 70
![Page 177: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/177.jpg)
Conclusions
Outline
1 Introduction
2 Motivation for use of particle filtering
3 The basic idea
4 Some particle filtering methods
5 Handling constant parameters
6 Rao-Blackwellization
7 Prediction and smoothing
8 Multiple particle filters
9 Convergence issues
10 Conclusions
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 68 / 70
![Page 178: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/178.jpg)
Conclusions
Conclusions
PF is a Monte Carlo-based methodology for sequential estimation ofdynamic (and static) signals (states)
The performance of the method depends on the number of particlesthat are used and the way how new particles are proposed
PF can be improved if some of the states are tracked by optimal(Kalman) filters
Convergence results exist
Efforts of generalizing PF include application of the philosophy whenthe noise distributions in the state and observation equations areunknown
PF is computationally intensive but special hardware can be built thatexploits the parallelizability of PF
Challenges of PF include tracking in high-dimensional state spaces
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 69 / 70
![Page 179: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/179.jpg)
Conclusions
Conclusions
PF is a Monte Carlo-based methodology for sequential estimation ofdynamic (and static) signals (states)
The performance of the method depends on the number of particlesthat are used and the way how new particles are proposed
PF can be improved if some of the states are tracked by optimal(Kalman) filters
Convergence results exist
Efforts of generalizing PF include application of the philosophy whenthe noise distributions in the state and observation equations areunknown
PF is computationally intensive but special hardware can be built thatexploits the parallelizability of PF
Challenges of PF include tracking in high-dimensional state spaces
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 69 / 70
![Page 180: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/180.jpg)
Conclusions
Conclusions
PF is a Monte Carlo-based methodology for sequential estimation ofdynamic (and static) signals (states)
The performance of the method depends on the number of particlesthat are used and the way how new particles are proposed
PF can be improved if some of the states are tracked by optimal(Kalman) filters
Convergence results exist
Efforts of generalizing PF include application of the philosophy whenthe noise distributions in the state and observation equations areunknown
PF is computationally intensive but special hardware can be built thatexploits the parallelizability of PF
Challenges of PF include tracking in high-dimensional state spaces
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 69 / 70
![Page 181: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/181.jpg)
Conclusions
Conclusions
PF is a Monte Carlo-based methodology for sequential estimation ofdynamic (and static) signals (states)
The performance of the method depends on the number of particlesthat are used and the way how new particles are proposed
PF can be improved if some of the states are tracked by optimal(Kalman) filters
Convergence results exist
Efforts of generalizing PF include application of the philosophy whenthe noise distributions in the state and observation equations areunknown
PF is computationally intensive but special hardware can be built thatexploits the parallelizability of PF
Challenges of PF include tracking in high-dimensional state spaces
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 69 / 70
![Page 182: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/182.jpg)
Conclusions
Conclusions
PF is a Monte Carlo-based methodology for sequential estimation ofdynamic (and static) signals (states)
The performance of the method depends on the number of particlesthat are used and the way how new particles are proposed
PF can be improved if some of the states are tracked by optimal(Kalman) filters
Convergence results exist
Efforts of generalizing PF include application of the philosophy whenthe noise distributions in the state and observation equations areunknown
PF is computationally intensive but special hardware can be built thatexploits the parallelizability of PF
Challenges of PF include tracking in high-dimensional state spaces
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 69 / 70
![Page 183: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/183.jpg)
Conclusions
Conclusions
PF is a Monte Carlo-based methodology for sequential estimation ofdynamic (and static) signals (states)
The performance of the method depends on the number of particlesthat are used and the way how new particles are proposed
PF can be improved if some of the states are tracked by optimal(Kalman) filters
Convergence results exist
Efforts of generalizing PF include application of the philosophy whenthe noise distributions in the state and observation equations areunknown
PF is computationally intensive but special hardware can be built thatexploits the parallelizability of PF
Challenges of PF include tracking in high-dimensional state spaces
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 69 / 70
![Page 184: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/184.jpg)
Conclusions
Conclusions
PF is a Monte Carlo-based methodology for sequential estimation ofdynamic (and static) signals (states)
The performance of the method depends on the number of particlesthat are used and the way how new particles are proposed
PF can be improved if some of the states are tracked by optimal(Kalman) filters
Convergence results exist
Efforts of generalizing PF include application of the philosophy whenthe noise distributions in the state and observation equations areunknown
PF is computationally intensive but special hardware can be built thatexploits the parallelizability of PF
Challenges of PF include tracking in high-dimensional state spaces
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 69 / 70
![Page 185: The Particle Filtering Methodology in Signal · PDF fileThe Particle Filtering Methodology in Signal Processing Petar M. Djuri c Department of Electrical & Computer Engineering Stony](https://reader031.vdocument.in/reader031/viewer/2022030400/5a7082a07f8b9aa2538c23a9/html5/thumbnails/185.jpg)
Conclusions
The Particle Filtering Methodology in Signal Processing
Petar M. Djuric
Department of Electrical & Computer EngineeringStony Brook University
September 25, 2009
The Particle Filtering Methodology in Signal Processing Petar M. Djuric 70 / 70