Large scale visual navigation and community map building

Humans are able to navigate large, complex environments using two primary senses: vision and inertial (from the inner ear) The system presented in this thesis aims to provide a similar capability for robots and autonomous vehicles. By fusing visual and inertial data, we are able to track our motion and reconstruct the environment through which we move. The system presented here substantially advances the state of the art both in terms of the robustness of the estimate to real-world variability and the scale at which it works. This information is used to build a map of the environment as we move through it. A novel topological map construction is developed, which provides the basis for loop closing, map fusion, and local map updates in a scalable manner. This lays the foundation for "community map building", the ability for multiple agents to work together to generate, update, and maintain a map which may or may not be globally consistent. We also present Corvis, a software framework for high data-rate real-time data collection, processing, and visualization. We demonstrate our system on indoor and outdoor datasets as large as the city scale.