Comparison of state marginalization techniques in visual inertial navigation filters

Thalagala, Ravindu G. (2019) Comparison of state marginalization techniques in visual inertial navigation filters. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (4MB)

Abstract

The main focus of this thesis is finding and validating an efficient visual inertial navigation system (VINS) algorithm for applications in micro aerial vehicles (MAV). A typical VINS for a MAV consists of a low-cost micro electro mechanical system (MEMS) inertial measurement unit (IMU) and a monocular camera, which provides a minimum payload sensor setup. This setup is highly desirable for navigation of MAVs because highly resource constrains in the platform. However, bias and noise of lowcost IMUs demand sufficiently accurate VINS algorithms. Accurate VINS algorithms has been developed over the past decade but they demand higher computational resources. Therefore, resource limited MAVs demand computationally efficient VINS algorithms. This thesis considers the following computational cost elements in the VINS algorithm: feature tracking front-end, state marginalization technique and the complexity of the algorithm formulation. In this thesis three state-of-the-art feature tracking front ends were compared in terms of accuracy. (VINS-Mono front-end, MSCKF-Mono feature tracker and Matlab based feature tracker). Four state-ofthe- art state marginalization techniques (MSCKF-Generic marginalization, MSCKFMono marginalization, MSCKF-Two way marginalization and Two keyframe based epipolar constraint marginalization) were compared in terms of accuracy and efficiency. The complexity of the VINS algorithm formulation has also been compared using the filter execution time. The research study then presents the comparative analysis of the algorithms using a publicly available MAV benchmark datasets. Based on the results, an efficient VINS algorithm is proposed which is suitable for MAVs.

Actions (login required)

View Item