## A Probabilistic Approach for the Adaptive Integration of Multiple Visual Cues Using an Agent Framework (2002)

Citations: | 4 - 3 self |

### BibTeX

@TECHREPORT{Soto02aprobabilistic,

author = {Alvaro Soto and John Dolan and Owen Carmichael and Peng Chang and Sal Desiano and Soshi Iba and Pragyana Mishra},

title = {A Probabilistic Approach for the Adaptive Integration of Multiple Visual Cues Using an Agent Framework},

institution = {},

year = {2002}

}

### OpenURL

### Abstract

Most current machine vision systems suffer from a lack of flexibility to account for the high variability of unstructured environments. As the state of the world evolves, the potential knowledge provided by different visual attributes can change, breaking the initial assumptions of a non-adaptive vision system. This thesis develops a new comprehensive computational framework for the adaptive integration of information from different visual algorithms.

### Citations

7556 |
Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference
- Pearl
- 1988
(Show Context)
Citation Context ...tages of the framework presented here. 2.1 Probabilistic Inference through Bayesian Reasoning 2.1.1 Bayes’ Rule Bayesian theory provides a solid mathematical framework for reasoning under uncertaint=-=y [57]-=-. Using the language of probability theory, a Bayesian approach provides mechanisms to combine information in order to reason about different hypothetical 13ssolutions to a problem. The basic idea is ... |

3310 | Snakes: Active contour models
- Kass, Witkin, et al.
- 1988
(Show Context)
Citation Context ...ed to this area obtaining an edge image, as shown in Figure 3.8(b). Starting from a configuration similar to the one presented in Figure 3.8(c), a snake algorithm fits a curve to the edge information =-=[36]-=-, as shown in Figure 3.8(d). Using the resulting snake points a uniform B-spline curve [2] is fitted to the target boundary using 40 control points. After scaling these control points to a uniform siz... |

2966 |
Introduction to Statistical Pattern Recognition
- Fukunaga
- 1990
(Show Context)
Citation Context ... σ 2 k represents the variance in the k dimension calculated with respect to the MAP estimator. Thus, the UD index is equivalent to the trace of the covariance matrix or to the sum of its eigenvalues=-= [23]-=-. The second index is intended to evaluate the quality of the information provided by each agent in terms of uncertainty reduction. The intuition behind it is that, if an agent is providing good infor... |

2700 |
Density Estimation for Statistics and Data Analysis
- Silverman
- 1986
(Show Context)
Citation Context ...d statistics provides several techniques that can be used to estimate these pdfs from data. These techniques are mainly classified as parametric, semiparametric, and non-parametric estimation methods =-=[5, 69]. -=-1 This thesis refers to P (ht/et−1) as the dynamic prior or just as the prior distribution. 15sParametric Density Estimation The parametric approach assumes that the true pdf can be modeled by a clo... |

2180 | An iterative image registration technique with an application to stereo vision
- Lucas, Kanade
- 1981
(Show Context)
Citation Context ...n the optical flow vectors of the pixels inside each hypothesis in the sample set. The agent calculates the optical flow vectors using an implementation of the Lucas and Kanade optical flow algorithm =-=[44]-=-. This algorithm uses an image registration technique to find the flow vectors. To represent a likelihood function of the observations, the motion agent applies a Gaussian function to the average valu... |

1199 |
Novel Approach to Nonlinear/Non-Gaussian Bayesian State Estimation
- Gordon, Salmond, et al.
- 1993
(Show Context)
Citation Context ... Filter This section presents the particle filter algorithm from a Bayesian perspective. This view of the filter, which is commonly used in the computer vision literature, was originally presented in =-=[27]-=- as the bootstrap filter and then introduced to the computer vision community in [32] as the condensation algorithm. Chapter 5 presents an alternative and more general definition of the particle filte... |

650 | Real-Time Tracking of Non-Rigid Objects using Mean Shift
- Comaniciu, Ramesh, et al.
- 2000
(Show Context)
Citation Context ...e behaviors. The chapter exemplifies the main features of the framework by applying the methodology to visual information coming from specially designed computer simulations. The mean shift algorithm =-=[9]-=-, currently one of the most successful tracking tools, is tested in some of the simulations to illustrate the advantages of the framework presented here. 2.1 Probabilistic Inference through Bayesian R... |

607 |
Theory of Statistics
- Schervish
- 1995
(Show Context)
Citation Context ...e goal of the estimation process is to find adequate values for the parameters of the function. Usually, the parameters can be efficiently estimated using training data and maximum likelihood methods =-=[12]. -=-For the case of Bayesian inference, the analytical functions are commonly chosen from conjugate families, which provides a closed-form implementation of the multiplicative step of Bayes’ rule. The m... |

594 | Contour tracking by stochastic propagation of conditional density
- Isard, Blake
- 1996
(Show Context)
Citation Context ...ve. This view of the filter, which is commonly used in the computer vision literature, was originally presented in [27] as the bootstrap filter and then introduced to the computer vision community in =-=[32]-=- as the condensation algorithm. Chapter 5 presents an alternative and more general definition of the particle filter in terms of the statistical principle of importance sampling [17]. In that chapter,... |

573 | Filtering via simulation: auxiliary particle filter
- Pitt, Shephard
- 1999
(Show Context)
Citation Context ...duce the computational burden by collapsing redundant components in the mixture, the complexity of the resulting filter is still far from O(N). Pitt and Shephard propose the auxiliary particle filter =-=[58]-=-. They argue that the computational complexity of the particle filter when using an arbitrary importance function can be reduced by performing the sampling step in a higher dimension. To achieve this,... |

572 |
Animate vision
- Ballard
- 1991
(Show Context)
Citation Context ...tarts providing useful information to find more adequate appearance models. The idea of detectors and specialists is closely related to the attentional mechanisms studied in the area of active vision =-=[78, 1]-=-. Motivated by their biological counterparts, the computer vision community has increased its interest in the incorporation of attentional mechanisms. These attention mechanisms are usually seen as te... |

516 |
Statistical Inference
- Casella, Berger
- 1990
(Show Context)
Citation Context ...d statistics provides several techniques that can be used to estimate these pdfs from data. These techniques are mainly classified as parametric, semiparametric, and non-parametric estimation methods =-=[5, 69]. -=-1 This thesis refers to P (ht/et−1) as the dynamic prior or just as the prior distribution. 15sParametric Density Estimation The parametric approach assumes that the true pdf can be modeled by a clo... |

455 |
Monte Carlo Strategies in Scientific Computing
- Liu
- 2001
(Show Context)
Citation Context ...In fact, one of the most attractive properties of MC integration, and MC techniques in general, is their unbiased convergence at a rate of O(N −1/2 ), which is independent of the number of dimension=-=s [42]. -=-In contrast, the convergence rate of numerical grid-based methods is O(N −1/d ), which rapidly decreases with the number of dimensions [42]. In order to apply MC techniques, the challenge is to find... |

370 |
Bayesian Inference in Econometric Models Using Monte Carlo Integration.Econometrica
- Geweke
- 1989
(Show Context)
Citation Context ...by: J(y) ≈ 1 n n� f(y, xi) (5.2) i=1 where x1, ..., xn are iid samples from p(x). Using the law of large numbers, it is possible to show that this estimate asymptotically converges to the right va=-=lue [25]-=-. 71sThe advantages of MC techniques over deterministic numerical methods can be appreciated by analyzing a numerical example. Consider the problem of calculating the expected value of the following p... |

337 |
The estimation of the gradient of a density function, with application in pattern recognition
- Fukunaga, Hostetler
- 1975
(Show Context)
Citation Context ...r feedback between the different vision modules. Recently, Comaniciu et al. [9] present an application of the mean shift algorithm to real-time tracking of non-rigid objects. The mean shift procedure =-=[24]-=- is an adaptive local steepest ascent technique for climbing probability density functions. The idea behind the mean shift procedure is that, over small regions, the average shift about a center posit... |

331 | The Hearsay-II Speech-Understanding System: Integrating Knowledge to Resolve Uncertainty
- Erman, H-Roth, et al.
- 1980
(Show Context)
Citation Context ... sources. Using ideas proposed independently by Newell [52] and Simon [70], Reddy and Erman implemented the first blackboard system as part of the HEARSAY and HEARSAY II speech understanding programs =-=[63, 19]-=-. A blackboard model consists of three major components: the knowledge sources, the blackboard, and the control unit. A blackboard model divides a problem into knowledge sources, which are kept separa... |

301 | Monte Carlo localization: efficient position estimation for mobile robots
- Fox, Burgard, et al.
(Show Context)
Citation Context ...y wasted samples located in irrelevant parts of the true distribution 4 . Previous works to adaptively determine an adequate number of particles have failed to consider these two factors together. In =-=[21]-=- and [37] the likelihood of the observations is used to adaptively select the number of particles. The idea is to use the sum of non-normalized importance weights as a measure of the fit between sampl... |

288 |
An Introduction to Splines for Use in Computer Graphics and Geometric Modeling
- BARTELS, BEATTY, et al.
- 1987
(Show Context)
Citation Context ...ration similar to the one presented in Figure 3.8(c), a snake algorithm fits a curve to the edge information [36], as shown in Figure 3.8(d). Using the resulting snake points a uniform B-spline curve =-=[2]-=- is fitted to the target boundary using 40 control points. After scaling these control points to a uniform size, they are used to generate a continuous close contour around the silhouette of the targe... |

284 |
Visual Routines
- Ullman
- 1984
(Show Context)
Citation Context ...r interaction between the different visual modules. Perceptual vision is just a bottom-up datadriven process. Recently, several researchers have criticized this idea and proposed an alternative model =-=[75, 79]-=-, which considers perception as a distributed collection of task-specific, task-driven visual routines with strong feedback among the visual modules. For example, Ullman [75] describes a biologically ... |

276 | ICondensation: Unifying Low-Level and High-Level Tracking in a Stochastic Framework
- Isard, Blake
- 1998
(Show Context)
Citation Context ...to determine recognition policies using different visual modules. Darrel et al. [11] present a tracking system that combines information from stereo, color, and face pattern-matching. Isard and Blake =-=[32, 33]-=- propose a probabilistic approach for target tracking under a Bayesian framework. Sherrah and Gong [68] propose the use of covariance estimation to track pose and face position fusing skin color infor... |

257 |
Neural Network Perception for Mobile Robot Guidance
- Pomerleau
- 1993
(Show Context)
Citation Context ... modify the amount of information of the different visual attributes. As a consequence, the most adequate set of attributes to complete a given task is highly variable. As an example, consider ALVINN =-=[59]-=-, a perceptual visual system designed to steer a car in natural environments using a neural net learning algorithm. After training, the main internal features learned by ALVINN were the edges of the r... |

247 | Machine learning research: Four current directions
- Dietterich
- 1998
(Show Context)
Citation Context ...e, it is possible to find a related idea in the context of ensemble of classifiers. An ensemble of classifiers is a set of classifiers whose individual decisions are combined to classify new examples =-=[14]-=-. Each classifier can be considered as a different source of knowledge. Adaptation mechanisms are included in the policy used to combine the outputs of the individual classifiers. This kind of techniq... |

226 | On Sequential Simulation-Based Methods for Bayesian Filtering
- Doucet
- 1998
(Show Context)
Citation Context ...prior and the likelihood function. The interpretation of the particle filter in terms of importance sampling suggests that alternative distributions can be used to obtain the samples. For example, in =-=[16], -=-84 b)sDoucet shows that the importance function p(xt/xt−1, yt) minimizes the variance of the importance weights conditional upon the state trajectory and the history of observations. A low value for... |

211 | Moving target classification and tracking from realtime video
- Lipton, Fujiyoshi, et al.
- 1988
(Show Context)
Citation Context ...me way, recent successful vision systems to detect people or cars are examples of specific visual applications where good engineering work provides an off-line identification of key visual attributes =-=[11, 35, 40]-=-. Unfortunately, the previous strategy lacks the flexibility needed to account for the variability of most natural scenarios. In the case of a dynamic unconstrained environment, it is difficult to kno... |

205 |
Level Set Methods
- Sethian
- 1996
(Show Context)
Citation Context ...ntinuous close contour around the silhouette of the target, as shown in Figure 3.8(e). Finally, this close contour is used to generate the shape vector, which consists of the signed distance function =-=[66]-=- calculated from a regular grid of points inside the input image. Figure 3.8(f) shows a gray level image of the signed distance function obtained for the example, each point encodes the distance to th... |

204 |
A Vision of the Brain
- Zeki
- 1993
(Show Context)
Citation Context ...hey can find their way home using the visual changing size of relevant natural landmarks. In the case of the human visual system, people usually characterize objects using distinctive visual features =-=[79, 50]. Fo-=-r example, most people are comfortable with instructions such as “Go straight until you see the tall round building, then turn right and go until you see the red sign... ” . The previous examples ... |

190 |
Stochastic Models
- MAYBECK
- 1979
(Show Context)
Citation Context ...sian at all times, and the well-known Kalman Filter gives the optimal solution. In the case of nonlinear models it is possible to use the Extended Kalman Filter, but still under a Gaussian assumption =-=[48]-=-. Unfortunately, the parametric assumption severely limits the use of Bayesian inference for state estimation. High ambiguity is one of the inherent features that emerges in most unstructured environm... |

190 |
Tools for Statistical Inference
- Tanner
- 1996
(Show Context)
Citation Context ...king of the evolution of the state space. A mathematical justification of the approximation given by the particle filter can be stated in terms of factored sampling [28] and the method of composition =-=[72]-=-. Factored sampling provides a justification to approximate the posterior by a weighted set of particles, as expressed in Equation (2.7). The method of composition justifies that the particles obtaine... |

169 | An introduction to sequential Monte Carlo methods,” in Sequential Monte Carlo Methods in Practice
- Doucet, Freitas, et al.
- 2001
(Show Context)
Citation Context ...e filter as the basic stochastic sampling technique to perform Bayesian inference. This tool has been successfully applied in diverse engineering fields to solve a variety of problems in recent years =-=[17]-=-. The great success of the particle filter can be explained by its great efficiency and extreme versatility in approximating any functional non-linearity of the pdfs. The key idea is to represent a po... |

165 | Training models of shape from sets of examples
- Cootes, Cooper, et al.
- 1992
(Show Context)
Citation Context ...ght misalignment with respect to the shape model. The training examples are used to build a point distribution model using principal component analysis (PCA) in a similar way to the model proposed in =-=[10]-=-. After the PCA decomposition, the eigenvectors corresponding to the largest eigenvalues are considered to build the shape space. The algorithm chooses those eigenvalues that account for more than 95%... |

163 |
Simulation Approaches to General Probabilistic Inference on Belief Networks
- Shachter, Peot
- 1990
(Show Context)
Citation Context ...ximated 17susing a set of regularly spaced samples. The previous idea corresponds to the likelihood weighting algorithm, which is a stochastic sampling technique widely used for stochastic simulation =-=[67]-=-. In stochastic sampling a pdf is represented through a set of samples, known as particles, each with an associated weight representing its probability. The likelihood weighting algorithm can be easil... |

154 |
Integrated person tracking using stereo, color, and pattern detection
- Darrell, Gordon, et al.
- 1998
(Show Context)
Citation Context ...me way, recent successful vision systems to detect people or cars are examples of specific visual applications where good engineering work provides an off-line identification of key visual attributes =-=[11, 35, 40]-=-. Unfortunately, the previous strategy lacks the flexibility needed to account for the variability of most natural scenarios. In the case of a dynamic unconstrained environment, it is difficult to kno... |

151 |
Data fusion for sensory information processing systems
- Clark, Yuille
- 1990
(Show Context)
Citation Context ...to build posterior probability densities, which are used to draw all the inferences about relevant hypotheses. This is closely related to the theory of regularization commonly used in computer vision =-=[6]-=-. Regularization theory calls for minimizing a cost function made up of a data and a regularizer term. The data term considers the evidence while the regularizer term constrains the set of possible hy... |

142 | Direct methods for recovering motion
- Horn, Jr
- 1988
(Show Context)
Citation Context ... There have been several attempts to achieve robot navigation using depth information from the environment. Although some of these attempts use visual cues such as focusing [54] and depth from motion =-=[30]-=-, binocular stereovision has been the favorite method [39, 47]. Ratler and Nomad at CMU, and Robby and HMMWV at JPL are examples of mobile robots that use a stereovision system as the main sensor for ... |

141 |
Stereo Vision for Planetary Rovers: Stochastic Modeling to Near Real-Time Implementation
- Matthies
- 1992
(Show Context)
Citation Context ...ion using depth information from the environment. Although some of these attempts use visual cues such as focusing [54] and depth from motion [30], binocular stereovision has been the favorite method =-=[39, 47]-=-. Ratler and Nomad at CMU, and Robby and HMMWV at JPL are examples of mobile robots that use a stereovision system as the main sensor for navigation. In these systems the goal has been a geometrical r... |

135 |
A decision theoretic generalization of on-line learning and an application to boosting
- Freund, Schapire
- 1995
(Show Context)
Citation Context ...ce over the individual classifiers that make them up. There are several algorithms available to implement the ensemble of classifiers; among the most relevant are Mixture of Experts [76] and AdaBoost =-=[22]-=-. The work presented here differs in many ways with respect to the current algorithms used to build an ensemble of classifiers. One of the main differences resides in the type of interaction among the... |

132 | Probabilistic data association methods for tracking complex visual objects
- Rasmussen, Hager
- 2001
(Show Context)
Citation Context ...es of visual information by integrating and adaptively switching in and out visual algorithms. Using a modified version of the joint probabilistic data association filter (JPDAF), Rasmussen and Hager =-=[61, 60, 62]-=- present an adaptive system that integrates visual information to perform tracking tasks. The JPDAF is an extension of the Kalman filter to deal with the data association problem in a Bayesian framewo... |

100 | A direct demonstration of functional specialization in the human visual cortex - Zeki, Watson, et al. - 1991 |

99 | Metropolized independent sampling with comparaisons to rejection sampling and importance sampling
- Liu
- 1996
(Show Context)
Citation Context ...placing (5.12) in (5.9) we obtain: N = σ 2 IS ≈ V arp(x) (V arq(w) + 1) (5.12) NIS V arq(w) + 1 = NIS Eq(w 2 ) ≈ NIS �N i=1 wi �N i=1 w2 i (5.13) Equation (5.13) coincides with a result by Li=-=u et al. [41]. Th-=-ey called this the N effective or the ”rule of thumb” of particle filters, which is used to quantify the degeneracy problem of particle sets. Including Equation (5.13) in (5.6), an alternative exp... |

79 |
Approximating posterior distributions by mixtures
- West
- 1993
(Show Context)
Citation Context ...es where the proposal density does not closely match the target distribution. The prior editing step suggested by Gordon in his seminal paper is an example of an acceptance/rejection test [27]. 85sIn =-=[77]-=-, West presents a method to adaptively build a suitable importance function to estimate a posterior distribution using a kernel-based approximation (mixture approximation). The basic idea is to apply ... |

69 | Cooperative Multisensor Video Surveillance
- Kanade
- 1997
(Show Context)
Citation Context ...me way, recent successful vision systems to detect people or cars are examples of specific visual applications where good engineering work provides an off-line identification of key visual attributes =-=[11, 35, 40]-=-. Unfortunately, the previous strategy lacks the flexibility needed to account for the variability of most natural scenarios. In the case of a dynamic unconstrained environment, it is difficult to kno... |

68 | KLD-Sampling: Adaptive Particle Filters
- Fox
(Show Context)
Citation Context ...s the quality of the match between the proposed and the true distribution, but it ignores the other important factor, i.e., the complexity of the true density. At NIPS’01, Fox introduces KLD-Samplin=-=g [20]-=-, a novel technique to adaptively estimate the number of particles that bound the error in the estimate of the true posterior. The error is measured by the Kullback-Leibler divergence (KL-divergence) ... |

62 | Visually-guided obstacle avoidance in unstructured environments
- Lorigo, Brooks, et al.
- 1997
(Show Context)
Citation Context ...and environment constraints to simplify the detection of relevant visual structures using just low-level visual cues such as particular colors, textures, and shapes, among others. For example, Lorigo =-=[43]-=- describes a mobile robot that is able to navigate in particular indoor environments by detecting the texture properties of the floor. The main characteristic of these systems is their simplicity. Typ... |

61 | Monte Carlo localization with mixture proposal distribution
- Thrun, Fox, et al.
- 2000
(Show Context)
Citation Context ...hoosing the index k using g(k/�yt) ∝ p(yt/µ k t ), with µ k t being the mean or other likely value associated with the density p(xt/x k t−1). In the context of mobile robot localization, Thrun=-= et al. [73]-=- notice that in cases where the likelihood function consists of a high peak, meaning a low level of noise in the observations, the particle filter suffers a degradation of performance. This suggests t... |

54 | Camera calibration without feature extraction
- Robert
- 1995
(Show Context)
Citation Context ... dimensions and position with respect to the cameras. Using this information the agent calculates the intrinsic and extrinsic parameters of the cameras using the calibration method proposed by Robert =-=[64]-=-. The next step consists in finding the disparities between the projection of world points into the two images. To facilitate this task, the agent applies two pre-processing steps to the input images.... |

52 | Incremental focus of attention for robust visual tracking”, CVPR, p 189 –195
- Toyama, Hager
- 1996
(Show Context)
Citation Context ...oing reference, and as a positioning system. 4.5 Computer Vision and Target Tracking There have been several attempts to build visual systems for the case of dynamic target tracking. Toyama and Hager =-=[74]-=- present a target tracking architecture based on an incremental focus of attention. The system consists of different trackers organized in a top-down fashion, where the trackers in the top layers prov... |

47 | Insect navigation en route to the goal: multiple strategies for the use of landmarks
- Collett
- 1996
(Show Context)
Citation Context ...a natural environment. Insects are able to navigate around a 3D cluttered world, avoiding obstacles, recognizing paths, landing in many places, taking off to avoid capture, and so on. Several studies =-=[7, 8]-=- show that insects, such as bees, move around their environment using a set of visual skills based on simple visual 1scues such as color, position, orientation, and relative change of size from differ... |

37 | Classification and Regression using Mixtures of Experts
- Waterhouse
- 1997
(Show Context)
Citation Context ... improve performance over the individual classifiers that make them up. There are several algorithms available to implement the ensemble of classifiers; among the most relevant are Mixture of Experts =-=[76]-=- and AdaBoost [22]. The work presented here differs in many ways with respect to the current algorithms used to build an ensemble of classifiers. One of the main differences resides in the type of int... |

36 |
Some problems of the basic organization in problem-solving programs
- Newell
- 1962
(Show Context)
Citation Context ...telligence (AI) domain the blackboard model for problem solving is one of the first attempts to adaptively integrate different types of knowledge sources. Using ideas proposed independently by Newell =-=[52]-=- and Simon [70], Reddy and Erman implemented the first blackboard system as part of the HEARSAY and HEARSAY II speech understanding programs [63, 19]. A blackboard model consists of three major compon... |

36 | Joint probabilistic techniques for tracking multi-part objects
- Rasmussen, Hager
- 1998
(Show Context)
Citation Context ...es of visual information by integrating and adaptively switching in and out visual algorithms. Using a modified version of the joint probabilistic data association filter (JPDAF), Rasmussen and Hager =-=[61, 60, 62]-=- present an adaptive system that integrates visual information to perform tracking tasks. The JPDAF is an extension of the Kalman filter to deal with the data association problem in a Bayesian framewo... |