Practical methods for Gaussian mixture filtering and smoothing

Detta är en avhandling från Chalmers University of Technology

Sammanfattning: In many applications, there is an interest in systematically and sequentially estimating quantities of interest in a dynamical system, using indirect and inaccurate sensor observations. There are three important sub-problems of sequential estimation: prediction, filtering and smoothing. The objective in the prediction problem is to estimate the future states of the system, using the observations until the current point in time. In the filtering problem, we seek to estimate the current state of the system, using the same information and in the smoothing problem, the aim is to estimate a past state. The smoothing estimate has the advantage that it offers the best performance on average compared to filtering and prediction estimates. Often, the uncertainties regarding the system and the observations are modeled using Gaussian mixtures (GMs). The smoothing solutions to GMs are usually based on pruning approximations, which suffer from the degeneracy problem, resulting in inconsistent estimates. Solutions based on merging have not been explored well in the literature. We address the problem of GM smoothing using both pruning and merging approximations. We consider the two main smoothing strategies of forward-backward smoothing (FBS) and two-filter smoothing (TFS), and develop novel algorithms for GM smoothing which are specifically tailored for the two principles. The FBS strategy involves forward filtering followed by backward smoothing. The existing literature provides pruning-based solutions to the forward filtering and the backward smoothing steps involved. In this thesis, we present a novel solution to the backward smoothing step of FBS, when the forward filtering uses merging methods. The TFS method works by running two filtering steps: forward filtering and backward filtering. It is not possible to apply the pruning or merging strategies to the backward filtering, as it is not a density function. To the best of our knowledge, there does not exist practical approximation techniques to reduce the complexity of the backward filtering. Therefore, in this thesis we propose two novel techniques to approximate the output of the backward filtering, which we call intragroup approximation and smoothed posterior pruning. We also show that the smoothed posterior pruning technique is applicable to forward filtering as well. The FBS and TFS solutions based on the proposed ideas are implemented for a single target tracking scenario and are shown to have similar performance with respect to root mean squared error, normalized estimation error squared, computational complexity and track loss. Compared to the FBS based on N-scan pruning, both these algorithms provide estimates with high consistency and low complexity.