Vesa Klumpp

Dirac mixture approximation of multivariate Gaussian densities

For the optimal approximation of multivariate Gaussian densities by means of Dirac mixtures, i.e., by means of a sum of weighted Dirac distributions on a continuous domain, a novel systematic method is introduced. The parameters of this approximate density are calculated by minimizing a global distance measure, a generalization of the well-known Cramérvon Mises distance to the multivariate case. This generalization is obtained by defining an alternative to the classical cumulative distribution, the Localized Cumulative Distribution (LCD). In contrast to the cumulative distribution, the LCD is unique and symmetric even in the multivariate case. The resulting deterministic approximation of Gaussian densities by means of discrete samples provides the basis for new types of Gaussian filters for estimating the state of nonlinear dynamic systems from noisy measurements.

Localized Cumulative Distributions and a multivariate generalization of the Cramér-von Mises distance

This paper is concerned with distances for comparing multivariate random vectors with a special focus on the case that at least one of the random vectors is of discrete type, i.e., assumes values from a discrete set only. The first contribution is a new type of characterization of multivariate random quantities, the so called localized cumulative distribution (LCD) that, in contrast to the conventional definition of a cumulative distribution, is unique and symmetric. Based on the LCDs of the random vectors under consideration, the second contribution is the definition of generalized distance measures that are suitable for the multivariate case. These distances are used for both analysis and synthesis purposes. Analysis is concerned with assessing whether a given sample stems from a given continuous distribution. Synthesis is concerned with both density estimation, i.e., calculating a suitable continuous approximation of a given sample, and density discretization, i.e., approximation of a given continuous random vector by a discrete one.

A novel Bayesian method for fitting a circle to noisy points

This paper introduces a novel recursive Bayesian estimator for the center and radius of a circle based on noisy points. Each given point is assumed to be a noisy measurement of an unknown true point on the circle that is corrupted with known isotropic Gaussian noise. In contrast to existing approaches, the novel method does not make assumptions about the true points on the circle, where the measurements stem from. Closed-form expressions for the measurement update step are derived. Simulations show that the novel method outperforms standard Bayesian approaches for circle fitting.

Density trees for efficient nonlinear state estimation

In this paper, a new class of nonlinear Bayesian estimators based on a special space partitioning structure, generalized Octrees, is presented. This structure minimizes memory and calculation overhead. It is used as a container framework for a set of node functions that approximate a density piecewise. All necessary operations are derived in a very general way in order to allow for a great variety of Bayesian estimators. The presented estimators are especially well suited for multi-modal nonlinear estimation problems. The running time performance of the resulting estimators is first analyzed theoretically and then backed by means of simulations. All operations have a linear running time in the number of tree nodes.

Dirac mixture trees for fast suboptimal multi-dimensional density approximation

We consider the problem of approximating an arbitrary multi-dimensional probability density function by means of a Dirac mixture density. Instead of an optimal solution based on minimizing a global distance measure between the true density and its approximation, a fast suboptimal anytime procedure is proposed, which is based on sequentially partitioning the state space and component placement by local optimization. The proposed procedure adaptively covers the entire state space with a gradually increasing resolution. It can be efficiently implemented by means of a pre-allocated tree structure in a straightforward manner. The resulting computational complexity is linear in the number of components and linear in the number of dimensions. This allows a large number of components to be handled, which is especially useful in high-dimensional state spaces.

Optimal dirac approximation by exploiting independencies

The sample-based recursive prediction of discrete-time nonlinear stochastic dynamic systems requires a regular reapproximation of the Dirac mixture densities characterizing the state estimate with an exponentially increasing number of components. For that purpose, a systematic approximation method is proposed that is deterministic and guaranteed to minimize a new type distance measure, the so called modified Cramér-von Mises distance. A huge increase in approximation performance is achieved by exploiting structural independencies usually occurring between the random variables used as input to the system. The corresponding prediction step achieves optimal performance when no further assumptions can be made about the system function. In addition, the proposed approach shows a much better convergence compared to the prediction step of the particle filter and by far fewer Dirac components are required for achieving a given approximation quality. As a result, the new approximation method opens the way for the development of new fully deterministic and optimal stochastic state estimators for nonlinear dynamic systems.

Bounding linearization errors with sets of densities in approximate Kalman filtering

Applying the Kalman filtering scheme to linearized system dynamics and observation models does in general not yield optimal state estimates. More precisely, inconsistent state estimates and covariance matrices are caused by neglected linearization errors. This paper introduces a concept for systematically predicting and updating bounds for the linearization errors within the Kalman filtering framework. To achieve this, an uncertain quantity is not characterized by a single probability density anymore, but rather by a set of densities and accordingly, the linear estimation framework is generalized in order to process sets of probability densities. By means of this generalization, the Kalman filter may then not only be applied to stochastic quantities, but also to unknown but bounded quantities. In order to improve the reliability of Kalman filtering results, the last-mentioned quantities are utilized to bound the typically neglected nonlinear parts of a linearized mapping.

Combined set-theoretic and stochastic estimation: A comparison of the SSI and the CS filter

In estimation theory, mainly set-theoretic or stochastic uncertainty is considered. In some cases, especially when some statistics of a distribution are not known or additional stochastic information is used in a set-theoretic estimator, both types of uncertainty have to be considered. In this paper, two estimators that cope with combined stochastic and set-theoretic uncertainty are compared, namely the Set-theoretic and Statistical Information filter, which represents the uncertainty by means of random sets, and the Credal State filter, in which the state information is given by sets of probability density functions. The different uncertainty assessment in both estimators leads to different estimation results, even when the prior information and the measurement and system models are equal. This paper explains these differences and states directions, when which estimator should be applied to a given estimation problem.

Direct fusion of Dirac mixture densities using an efficient approximation in joint state space

In this paper, we present a direct fusion algorithm for processing the combination of two Dirac mixture densities. The proposed approach allows the multiplication of two Dirac mixture densities without requiring identical support and thus enables the fusion of two independently generated sample sets. The resulting posterior Dirac mixture density is an approximation of the true continuous density that would result from the processing of the underlying true continuous density functions. This procedure is based on a suboptimal greedy approximation of the joint state space by means of a Dirac mixture that iteratively increases the resolution of the fusion result while considering only the relevant regions in the joint state space, where the fusion constraint holds.

Approximate Nonlinear Bayesian Estimation Based on Lower and Upper Densities

Recursive calculation of the probability density function characterizing the state estimate of a nonlinear stochastic dynamic system in general cannot be performed exactly, since the type of the density changes with every processing step and the complexity increases. Hence, an approximation of the true density is required. Instead of using a single complicated approximating density, this paper is concerned with bounding the true density from below and from above by means of two simple densities. This provides a kind of guaranteed estimator with respect to the underlying true density, which requires a mechanism for ordering densities. Here, a partial ordering with respect to the cumulative distributions is employed. Based on this partial ordering, a modified Bayesian filter step is proposed, which recursively propagates lower and upper density bounds. A specific implementation for piecewise linear densities with finite support is used for demonstrating the performance of the new approach in simulations