Data fusion for localization using state estimation and machine learning

Abstract

Using multiple sensory information is acknowledged as one of the major topics in the navigation of aerial and ground vehicles. This doctoral thesis considers localization as a state estimation problem, which is solved by data fusion techniques and supported by machine learning methods. It attempts to address the issue of developing a better localization system for a ground vehicle by seeking the best possible pose estimator (i.e., position, velocity and attitude) and improving its robustness to unexpected sensor measurements. The vehicle of interest is represented by a skid-steer tracked mobile robot; however, all the algorithms work with a sensory set, which can be with minor changes deployed on any vehicle, legged, wheeled, tracked, or aerial. First part of this thesis explores the development of different state estimation architectures, which exploit the extended Kalman filter for full 3D dead reckoning (i.e., incremental or relative pose estimation). The purpose of this part is to use inertial and odometry dead reckoning to its optimal extent, considering both the performance and computational complexity. Such combination of proprioceptive sensory modalities used on a ground vehicle is expected to provide the core localization— foundation for any other higher level localization or navigation systems. Second part of the dissertation investigates means of improving overall robustness and performance of the multi-modal state estimation. Different sensory modalities are prone to various types of errors, especially in an environment that changes dynamically. Therefore, the thesis shows the importance of identifying and rejecting unexpected or erroneous measurements. The multi-modal data fusion is based on inertial and odometry measurements aided by information from a camera and laser range finder. These two exteroceptive modalities are in particular prone to real-world disturbances, therefore, they are the subject of anomaly detection process. Various state-of-the-art machine learning methods (i.e., logistic regression, Support Vector Machines, Gaussian Mixture Models and Gaussian Processes) are applied in a Kalman filter framework to monitor the measurements and overcome the commonly used covariance monitoring and chi-squared gating test. Verification of all the techniques in this thesis is supported by extensive experimental datasets collected with a real mobile robot in both indoor and challenging outdoor environments.