## Better proposal distributions: Object tracking using unscented particle filter (2001)

### Cached

### Download Links

Citations: | 81 - 2 self |

### BibTeX

@INPROCEEDINGS{Rui01betterproposal,

author = {Yong Rui and Yunqiang Chen},

title = {Better proposal distributions: Object tracking using unscented particle filter},

booktitle = {},

year = {2001},

pages = {786--793}

}

### Years of Citing Articles

### OpenURL

### Abstract

Tracking objects involves the modeling of non-linear non-Gaussian systems. On one hand, variants of Kalman filters are limited by their Gaussian assumptions. On the other hand, conventional particle filter, e.g., CONDENSATION, uses transition prior as the proposal distribution. The transition prior does not take into account current observation data, and many particles can therefore be wasted in low likelihood area. To overcome these difficulties, unscented particle filter (UPF) has recently been proposed in the field of filtering theory. In this paper, we introduce the UPF framework into audio and visual tracking. The UPF uses the unscented Kalman filter to generate sophisticated proposal distributions that seamlessly integrate the current observation, thus greatly improving the tracking performance. To evaluate the efficacy of the UPF framework, we apply it in two real-world tracking applications. One is the audio-based speaker localization, and the other is the visionbased human tracking. The experimental results are compared against those of the widely used CONDENSATION approach and have demonstrated superior tracking performance. 1.

### Citations

1190 |
A Novel Approach to Non-Linear and Non-Gaussian Bayesian State Estimation
- Gordon, J, et al.
- 1993
(Show Context)
Citation Context ...ilters in recent years. The first appearance of particle filters can be traced back to 1950s [7]. While almost dormant in the seventies, there is a renaissance of this technique in the early nineties =-=[6,8,14,17]-=-, due to the massive increases in computing power. However, most of them use the state transition prior p(xt|xt-1) as the proposal distribution to draw particles from [8,18]. Because the state transit... |

593 | Contour tracking by stochastic propagation of conditional density
- ISARD, BLAKE
- 1996
(Show Context)
Citation Context ...have demonstrated superior tracking performance. 1. Introduction Reliable object tracking in complex audio-visual environment is an important task. Its applications include human computer interaction =-=[8,9]-=-, teleconferencing [19,20], and surveillance [12], among many others. It is also a very challenging task in that objects’ state space representation can be highly non-linear and the observation (e.g.,... |

510 | Sequential monte carlo methods for dynamic systems
- Liu, Chen
- 1998
(Show Context)
Citation Context ...ilters in recent years. The first appearance of particle filters can be traced back to 1950s [7]. While almost dormant in the seventies, there is a renaissance of this technique in the early nineties =-=[6,8,14,17]-=-, due to the massive increases in computing power. However, most of them use the state transition prior p(xt|xt-1) as the proposal distribution to draw particles from [8,18]. Because the state transit... |

272 | Icondensation: Unifying low-level and high-level tracking in a stochastic framework
- Isard, Blake
- 1998
(Show Context)
Citation Context .... They sample from the transition prior and calculate the importance weight as follows: ( i) ( i) ( i) ~ ( i) ~ ( i) p( yt | xt ) p( xt | xt−1) ~ ( i) ( i) wt = wt− 1 = w 1 ( | ) ( i) ( i) t− p yt xt =-=(9)-=- q( xt | x0: t−1, y1: t ) Even though simple to implement, this proposal results in higher Monte Carlo variance and thus worse performance [5,17]. Comparing the transition prior p(xt|xt-1) with the ge... |

225 | On sequential simulation-based methods for bayesian filtering
- Doucet
- 1998
(Show Context)
Citation Context ...om transition prior may have very low likelihood, and their contributions to the posterior estimation become negligible. This type of particle filters is prone to be distracted bysbackground clutters =-=[5,11,17]-=-. For clarity, in this paper, we refer this type of filters as the conventional particle filters. Inside the computer vision community, particle filters has also enjoyed considerable attention. Follow... |

202 |
Sequential Imputations and Bayesian Missing Data Problems
- Kong, Liu, et al.
- 1994
(Show Context)
Citation Context .... Introduction Reliable object tracking in complex audio-visual environment is an important task. Its applications include human computer interaction [8,9], teleconferencing [19,20], and surveillance =-=[12]-=-, among many others. It is also a very challenging task in that objects’ state space representation can be highly non-linear and the observation (e.g., audio and/or visual sensory data) is almost alwa... |

184 | Partitioned sampling, articulated objects, and interface-quality hand tracker
- MacCormick, Isard
- 2000
(Show Context)
Citation Context ...sion community, particle filters has also enjoyed considerable attention. Following the pioneering work of CONDENSATION [8], various improvements and extensions have been proposed for visual tracking =-=[2,9,16]-=-. Because the original CONDENSATION algorithm uses the state transition prior as its proposal distribution, it belongs to the conventional particle filters. To design better proposal distributions for... |

167 | 2001a), An introduction to sequential monte carlo methods
- Doucet, Freitas, et al.
(Show Context)
Citation Context ...0 : t N N→∞ ∑ i= 1 ( i) ( i) E ( g( x )) = lim g( x ) w ( x ) (4) Furthermore, as N tends to infinity, the posterior distribution p can be approximated by the properly weighted particles drawn from q =-=[4,14,17]-=-: 0: t N ( i) pˆ ( x0 : t | y1: t ) = ∑ w i= t ( x t i 1 0: ) ( ) x0: t t 0: t 0: t δ ( dx ) (5) There are two important points worth emphasizing here. First, the definition says that an unknown distr... |

122 | A general method for approximating nonlinear transformations of probability distributions
- Julier, Uhlmann
- 1996
(Show Context)
Citation Context ... filter (UKF) that can accurately compute the mean and covariance of y=g(x),whereg( ) is an arbitrary function, up to the second order (third in Gaussion prior) of the Taylor series expansion of g( ) =-=[10]-=-. While UKF is significantly better than EKF in density statistics estimation, it still assumes a Gaussian parametric form of the posterior, thus cannot handle multi-modal distributions. The non-param... |

114 | A probabilistic framework for matching temporal trajectories
- Black, Jepson
- 1998
(Show Context)
Citation Context ...sion community, particle filters has also enjoyed considerable attention. Following the pioneering work of CONDENSATION [8], various improvements and extensions have been proposed for visual tracking =-=[2,9,16]-=-. Because the original CONDENSATION algorithm uses the state transition prior as its proposal distribution, it belongs to the conventional particle filters. To design better proposal distributions for... |

100 | Learning to track the visual motion of contours
- Blake, Isard, et al.
(Show Context)
Citation Context ...nctions around the predicted values, extended Kalman filter (EKF) is proposed to solve non-linear system problems. It is first introduced in control theory [1] and later on applied in visual tracking =-=[3]-=-. Because of its first-order approximation of Taylor series expansion, EKF finds only limited success in tracking visual objects [8]. In recent years, Julier and Uhlmann develop an unscented Kalman fi... |

70 | Nonlinear filtering for speaker tracking in noisy and reverberant environments
- Vermaak, Blake
- 2000
(Show Context)
Citation Context ...or tracking performance. 1. Introduction Reliable object tracking in complex audio-visual environment is an important task. Its applications include human computer interaction [8,9], teleconferencing =-=[19,20]-=-, and surveillance [12], among many others. It is also a very challenging task in that objects’ state space representation can be highly non-linear and the observation (e.g., audio and/or visual senso... |

51 |
Voice source localization for automatic camera pointing system in video conferencing
- Wang, Chu
- 1997
(Show Context)
Citation Context ... Y [ X i, t| t−1 + K ( y − y x i, t| t−1 K = P t t| t−1 − y − x xt yt ), t| t−1 t| t−1 P −1 yt yt P = P t ][ Y ][ Y i, t| t−1 x i, t| t−1 t| t−1 − y − y T t| t−1 T t| t−1 − K P t ] ] yt yt K T t (19) =-=(20)-=- (21) (22) Compared with the EKF [1], the UKF does not need to explicitly calculate the Jacobians or Hessians. Therefore, the UKF not only outperforms the EKF in accuracy (second order approximation v... |

46 |
Optimal Filtering. Englewood Cliffs
- Anderson, Moore
- 1979
(Show Context)
Citation Context ...ise, respectively. If f( ) and h( ) are linear functions and if Gaussian distribution is assumed for xt, mt and nt , p(xt |xt-1, y0:t) has an analytical solution which is the well-known Kalman filter =-=[1]-=-. Unfortunately, tracking objects in real-world environment seldom satisfies Kalman filter’s requirements. For example, in human tracking, background clutter may resemble the human face, and in sound ... |

45 |
The Unscented Particle Filter
- Merwe, Doucet, et al.
- 2000
(Show Context)
Citation Context ...ilters in recent years. The first appearance of particle filters can be traced back to 1950s [7]. While almost dormant in the seventies, there is a renaissance of this technique in the early nineties =-=[6,8,14,17]-=-, due to the massive increases in computing power. However, most of them use the state transition prior p(xt|xt-1) as the proposal distribution to draw particles from [8,18]. Because the state transit... |

36 | Automating camera management for lecture room environments
- Liu, Rui, et al.
- 2001
(Show Context)
Citation Context ...an audio-data-based tracking system in Section 4 and a visual-data-based tracking system in Section 5. 4. UPF tracking using audio sensory data In many applications, including automated lecture rooms =-=[15]-=- and teleconferencing [19,20], we need to reliably track the location of the person who is talking. This is usually done by using a microphone array and a pan/tilt/zoom camera, as shown in Figure 1(a)... |

29 |
Poor man's Monte Carlo
- Hammersley, Morton
- 1954
(Show Context)
Citation Context ...osteriors can be estimated sequentially over time. This technique is more popularly known as the particle filters in recent years. The first appearance of particle filters can be traced back to 1950s =-=[7]-=-. While almost dormant in the seventies, there is a renaissance of this technique in the early nineties [6,8,14,17], due to the massive increases in computing power. However, most of them use the stat... |

27 | How does CONDENSATION behave with a finite number of samples
- King, Forsyth
- 2000
(Show Context)
Citation Context ...) P ) i = x ( m) 0 ( m) i = λ /( n + λ), = W ( m) i 2 λ = α ( n + κ ) − n x x x x x x x W ( c) 0 x i i i = 1, K, n i = n + 1, K, 2n = W = 1/( 2⋅ ( n + λ)) x ( m) 0 x 2 + ( 1− α + β ) x i = 1, K, 2n x =-=(11)-=- where κ is a scaling parameter that controls the distance between the sigma points and the mean x . α is a positive scaling parameter that controls the higher order effects resulted from the non-line... |

20 |
Simultaneous Tracking and Verification via Sequential Posterior Estimation
- Li, Chellappa
- 2000
(Show Context)
Citation Context ...igma points through the nonlinear transformation: Yi = g( X i ) i = 0, K2nx (12) 3. Compute the mean and covariance of y as follows: y = 2n x ∑ i= 0 W ( m i 2n x ∑ ) ( c) T Y , P = W ( Y − y)( Y − y) =-=(13)-=- i y i= 0 The mean and covariance of y is guaranteed to be accurate up to the second order of the Taylor series expansion. 3.2. The unscented Kalman filter The unscented Kalman filter (UKF) can be imp... |

1 | Particle filtering for non-stationary speech modeling and enhancement
- Vermaak, Andrieu, et al.
- 2000
(Show Context)
Citation Context ...ficult if not impossible. Instead, the conventional particle filters have chosen to trade the optimality with easy-implementation by using the transition prior p(xt|xt-1) as the proposal distribution =-=[6,8,18]-=-. They sample from the transition prior and calculate the importance weight as follows: ( i) ( i) ( i) ~ ( i) ~ ( i) p( yt | xt ) p( xt | xt−1) ~ ( i) ( i) wt = wt− 1 = w 1 ( | ) ( i) ( i) t− p yt xt ... |