Towards Robust Semantic Simultaneous Localization and Mapping in GPS-Denied Environments

Abstract

Simultaneous localization and mapping (SLAM) is the concurrent estimation of a robot's trajectory and a reconstruction of its surroundings. Accuracy and robustness of these estimates is of great significance, particularly in applications that concern safety and security. Yet, the estimation process is vulnerable to a wide range of noise patterns and error sources, that hinder the estimation accuracy and lead to severe artifacts in the reconstructed map and estimated trajectory. Examples of such sources include measurement noise, sensor noise, environmental challenges, or limitations of the employed optimization algorithms that incapacitates guaranteeing optimality of the optimization solutions. To that end, this thesis researches introducing learning approaches in the SLAM pipeline to tackle the challenge of enhancing estimation accuracy from different aspects. More specifically, the following three contributions are made: (1) a novel stacked long short term memory (LSTM) network is developed to learn to identify and reduce trajectory estimation errors resulting from several error sources, with an application to object-based semantic SLAM, (2) a novel pose graph neural classifier that is capable of classifying, with high prediction accuracy, 2D pose-graphs into optimal or non-optimal is established, (3) a novel graph neural network is developed to correct SLAM pose-graph estimates in order to achieve higher estimation accuracy, given an initial estimate of the trajectory. The proposed approaches are tested and evaluated using simulated and real-time experiments, including publicly available SLAM datasets. Each approach was compared against and proved to outperform its counterpart state-of-the-art solutions.

Date of Award	Dec 2020
Original language	American English