Mobility tracking algorithms for cellular networks based on Kalman filtering

Abstract

This thesis is an in depth theoretical and practical survey of dynamic mobility tracking systems specifically for cellular networks. A user mobility state model that is originally proposed for tracking targets in tactical weapons systems is discussed. This mobility model captures a large range of mobility by modeling acceleration(manuever) as driven by a discrete semi-Markovian command process and a Gaussian time-correlated random process. Linear and nonlinear observation models are presented. For nonlinear model, received signal strength indicator model of a cellular communication network is considered. Based on these models, tracking algorithms are presented and simulated in Matlab. Algorithms differ in the type of observation model they use(linear or nonlinear) and in the way they treat semi-Markovian manuever component and they employ variants of the Kalman filter namely: traditional Kalman filter, extended Kalman filter, uscented Kalman filter or adaptive Kalman filter. Linear observation command mode algorithm, nonlinear observation command mode algorithm and unscented Kalman filter based version of nonlinear observation command noise algorithm are introduced in this thesis. Matlab simulations and root mean square error statistics of simulations represented by sample mean and standard deviation of a number of root mean square error values show that treating command process as an additional state noise is a more accurate approach than treating it as a model variable within a mutliple model adaptive estimator. Extended and unscented Kalman filters are used to calculate predicted states and covariances when nonlinear observation model is used. Simulations and root mean square error statistics of simulations show that nonlinear structure of received signal strength indicator model causes large Gaussian approximation errors for posterior state probability and state and covariance prediction errors in extended and unscented Kalman filters, but smaller in uscented Kalman filter. Moreover, nonlinearity pronounces non-Gausianity of likelihood function used in the adaptive estimator as much as to terminate the estimator prematurely.