| Recent hardware developments have rendered controlled active vision a practical option for a broad range of problems: spanning applications as diverse as intelligent vehicle highway systems, robotic-assisted surgery, 3D reconstruction, inspection, vision assisted grasping, microassembly of micro-electromechanical systems (MEMS) and automated spacecraft docking. However, realizing this potential requires having a framework for synthesizing robust active vision systems, capable of moving beyond carefully controlled environments.; This thesis is a study of a novel approach to modeling and controlling of active vision systems. We show how recently developed robust identification techniques can be used to find a family of models for an active vision system in a unified sense, treating cameras, motors and image processing hardware as a single system without requiring any calibration.; Moreover, we use this family of models to design a robust controller by using μ-synthesis techniques. The designed controller is shown to be capable of performing satisfactorily in the presence of uncertainty, noise, changing camera parameters (various zoom values, from minimum zoom levels to almost maximum zoom levels), uncertain time delays (largely due to the time required by the image processing algorithm to locate the object in the scene), unmodeled dynamics, blurring of the image, etc.; These results are experimentally validated using a UniSight/BiSight robotic head-eye platform. |
