Appearance-Based Mobile Robot Localization and Map-Building in Unstructured Environments

Abstract

This thesis explains direct image processing of (circular) images from a catadioptric vision system. The image processing is then used in localization and map-building. Our research in vision-based navigation consists of localization that demonstrates the advantages of vision over other type of sensors. However, initial work is still limited by the assumption that a map of an environment is available. In particular, the assumption is often invalid due to dynamic changing of the surrounding appearance. Therefore, the simultaneous localization and map-building (SLAM) system is proposed. Our SLAM system is able to incrementally build a map and detect a loop in a cyclical environment. The system is able to produce a map with reasonable accuracy even when operating a robot in a large indoor area. Lastly, we present the integration of our appearance-based SLAM with other navigation tasks.