Models and Methods for Inertial and Visual Odometry

Abstract

Smartphones have become a part of everyday modern life. Their presence and capabilities have changed the way a lot of tasks are performed. Among them is navigation, how to find a way from one point to another. Satellite-based navigation allows for easy localization and path planning when traveling outdoors in the open. Another recent capability is augmented reality and local mapping. However, medium-scale indoor navigation and mapping have not been reliably solved. Challenges in these environments are characterized by an unreliable satellite signal and visual features which are not easily recognizable or not revisited. This thesis is focused on inertial and visual-inertial navigation for pedestrian dead reckoning. The systems, algorithms, datasets, and benchmarks proposed in this thesis are all designed with smartphones and pedestrians in mind. The central statement of the thesis is that careful consideration of uncertainty sources, the characteristics of pedestrian motion, and the quality of the available signals allows for a system that produces reliable odometry over medium-length sequences of pedestrian motion. Two sensor modalities are explored in this thesis, inertial and visual-inertial odometry. Pure inertial odometry relies on complementary signals to keep a reliable estimate of the pose. Multiple approaches to utilize these complementary signals are explored in this thesis. Visual-inertial odometry is approached with computational efficiency in mind. The measurements from the camera and the inertial sensors are integrated using a state space model driven by inertial odometry. This is to allow for efficient use of camera information and complementary online processing. This thesis also introduces a dataset which is a representative sample of the kind of use cases that are problematic for most available systems. This dataset is used in both modalities to train, evaluate and benchmark results in pedestrian dead reckoning. The main contributions of this thesis are the proposed algorithms and data. This thesis shows that multiple modalities of information and multiple approaches to state estimation can be used to complement each other. The complementary use of sequential Bayesian estimation and deep learning has been explored further by the community, and many opportunities to improve on it are still to be discovered.