Single point positioning by means of particle filtering on the GPU

High demands on accuracy and reliability of real-time navigation and positioning applications require the exploitation of multiple sensors. However, in many cases, the relations between the state vector and the observation space are of non-linear mathematical nature or observations have stochastic properties which deviate from those of Gaussian normal distributions. Thus, the extended Kalman filter and its variants are not always the most suitable choice. For such problems a particle filter, or more generally, sequential Monte Carlo methods, can increase the reliability of the estimates. Since particle filters are well-suited for parallel data processing, it will be shown how single point positioning GNSS solutions can be obtained when using a graphics processing unit (GPU) as a massive parallel computing device. The implementation of this approach and evaluation of its performance concerning real-time capability will be discussed as well as its precision and accuracy compared to a standard Kalman filter solution.

I. Introduction

Major advances in engineering and basic science have led to significant improvements of positioning, navigation and timing (PNT) over the last century and thus enabled a manifold of applications which are part of our daily life. Driven by the requirements of autonomous navigation, there are new challenges that need to be addressed in order to realize novel concepts for aircraft, ship and land vehicle navigation. Together with those challenges rises the question on how to determine the “most likely” state of a navigation system under the consideration of stochastic properties of the observations as well as restrictions in the solution space. Thus, new, improved and flexible methods for parameter estimation in dynamic systems are necessary.