Scientists at Defence R&D Canada - Suffield have been investigating autonomous operation of Unmanned Ground Vehicles (UGV) in complex indoor and outdoor environments. The ability of the UGV to navigate autonomously is largely dependent on the accuracy and robustness of its perception system which seeks to create an accurate model of the UGV's environment. One of the factors determining the requirement for a UGV's perception system is its operating environment. Designing a perception system capable of dealing with all types of environments is unfeasible at this time. In order to constrain the problem, many state of the art UGVs are designed from the ground up to operate within an assumed environment. If these assumptions are valid, the UGV can operate effectively, however, failure will occur when these environmental assumptions are incorrect. In order to alleviate this problem, it is desirable to have a perception system which can adapt to its changing environment. In order to do this, the UGV must understand the context of its environment and recognize when that context changes. One possible method of doing this is through the classification of video imagery.

This thesis proposes a perception system which adapts to its changing environment at run time (indoor/outdoor) through the use of vision and learning techniques. The system uses a digital elevation map when operating outdoors and relies on accurate GPS and IMU sensory data for its pose. When transitioning to indoor environments, the perception system automatically adjusts (via a classification algorithm) to use an Extended Kalman Filter based Simultaneous Localization and Mapping algoirthm. Numerous image feautres and learning techniques are analyzed to determine their suitability to indoor/outdoor classification. In addition, the accuracy and deficiencies of the mapping algorithms are also detailed.