high performance computing on graphics processing units: hgpu.org

In the field of Tracking and Data Fusion most, if not all, computations executed by a computer are carried out serially. The sole part of the process that is not entirely serial is the collection of data from multiple sensors, which can be executed in parallel. However, once the data is to be filtered the most likely candidate is a serial algorithm. This is due in large part to the algorithms themselves that have been developed over the last several decades for use on conventional computers that have been left void of parallel computing capabilities, until now. With the arrival of graphical processing units, or GPUs, the tracking community is in a favourable position to exploit the functionality of parallel processing in order to track a growing number of targets. The problem, however, begins with the sheer labour of having to convert all the pre-existing serial tracking algorithms into parallel ones. This is clearly a daunting task when one considers the extent to which the tracking community has gone to develop modern day filters such as Alpha Beta filters, Probabilistic Data Association filters, Interacting Multiple Model filters, and several dozen, if not hundred, variants of the aforementioned. It is most likely that these filters will find some kind of a parallelization in the near future as ever more sensors are dispersed throughout society and even more targets are monitored with these sensors. The volume of targets then becomes simply too unmanageable for a serial algorithm and more focus is placed iv on parallel ones. Yet, before the parallel algorithms can be utilized they have to be derived. It is the derivation of these parallel algorithms which is the focus of this thesis. However, it should be made clear that it would be impossible to formulate a parallelization for every filter found in the literature, and so the goal here is to direct the attention onto one filter in particular, the Kalman filter.