(A) study on robust particle filtering in terrain-referenced navigation

Abstract

This dissertation addresses a robust particle sampling strategy in terrain-referenced navigation problem. The study proposes an auxiliary sampled Gaussian mixture proposal distribution particle filter and a method of recursively instantiating multiple particle filter objects at the sigma points of the overall distribution. The proposal distribution of each particle filter stochastically nudges every particle towards the latest measurement, exhibiting a contracting feature of particles. Especially, the multimodal noise characteristic of radar altimeter measurement is remedied via a superposition of mode-associated stochastic nudging. The effect is reinforced by adapting the principle of auxiliary variable sampling. At the same time, operating multiple instances of particle filter at the sigma points force diversify the particle occupation over the state space. Particularly, the moments-preserving mixture transformation of the Gaussian distribution alleviates the ambiguity of determining variance each mixture as well as the use of heuristic parameters required in the general mixture particle filtering. The proposed method, which should be the combination of the two rationales, shows complementary feature via contraction-relaxation mechanism and becomes robust against potential degeneration and impoverishment of the particle filter and against large initial errors. The study focuses on the causes of the degeneration and impoverishment including multimodal measurement noise of terrain sensor, biased feature of the inertial navigation system, and severely nonlinear characteristic of terrain elevation. The extensive numerical experiments carried out in a various test scenarios shows the integrity and robustness of the proposed method.