Gabor-gist Visual Homing

Abstract

Many robotic systems are required to navigate or home to learned location using minimal resources. Autonomous robots are generally limited in computation and storage resources, imposing significant challenges on algorithm design. Particularly when only visual data is used, these algorithms need to be robust and efficient. In addition, independence from a scene model is preferred. Extraction of models and calibration procedures are time consuming and sensitive to changes in the environ- ment. Visual homing without a geometric model is studied in mapless or qualitative visual homing. In this thesis, we adopt a framework based on View-Sequenced Route Representation (VSRR) and contribute in two areas: Compact representa- tion of the path and visual homing along a desired route using the representation, and secondly develop an algorithm which localizes the robot using a novel concept we call eigensegments. The effectiveness of the system is demonstrated with both indoor and outdoor environments.