PARALLELIZED SIGMA POINT AND PARTICLE FILTERS FOR NAVIGATION PROBLEMS

摘要

Advanced filters like the sigma point and particle filters are more accurate than the extended Kalman filter for nonlinear and non-Gaussian applications, but experience drawbacks such as being computationally expensive with a serial implementation. However, since the majority of the computation can be carried out simultaneously, these filters are inherently well suited for parallel computing. This research leverages inexpensive and personal-level parallel computing architectures, such as the NVIDIA Graphics Processing Units (GPUs) and multi-core CPUs to exploit such parallelism. In particular, parallelization of the Unscented Kalman filter (UKF) and the bootstrap Particle Filter (PF) applied to an orbit determination problem and a loosely coupled GPS/INS integration problem is the main objective of this work. This work demonstrates that the UKF and the PF can be applied to the two navigation problems with great benefits in the presence of nonlinearities and non-Gaussian noises. An 8-time speedup is achieved for both filters with an 8-thread CPU, and up to two orders of magnitude speedups are achieved using a M2090 GPU. The results show that the two UKF implementations can be executed in real time without parallelization, but the two PF implementations can be executed in real time only when parallelized on a GPU.