Visual servoing means the control of actions from visual perceptions to achieve a given goal. For example the insertion of a thread into a needle's hole is made much easier and more robust by the precision and flexibility of the human visual control paradigm. The artifical implementation of this is challenging and current systems are still highly inflexible. In particular the requirement for calibrated cameras and robots prevents the widespread use of visual servo control in industry. However this would be highly benefical, allowing human labour to be replaced in dangerous or difficult to access environments, and ensuring competitiveness through efficient automation.
The aim of this project is to advance in uncalibrated visual control by the comprehensive integration of projective geometry, the mathematical formalism to model uncalibrated cameras. In contrast to other approaches, the proposed representation in projective space allows for view-point independent 3D goal definition, generation of global, convergent spatial trajectories, accurate Euclidean alignment by constraining projective features (visual cues) and the transfer of the visual information to arbitrary views. for subsequent efficient visual control at a high speed. In addition, quantitative Euclidean goals such as orthogonality, angular alignment or accuracy requirements will be expressed by means of visual cues in the projective domain. These is done by exploiting minimal knowledge about Euclidean structures in the scene or observing rigid robot motions. Recently available techniques such as view-transfer and projective reconstruction allow for the first time a theoretical study of view-points that are optimal with respect to certain performance criteria or that constitute a degenerate configuration. Crucial to the feasibility of the proposed approach is the robust estimation of the robot-to-image Jacobian and the solution of the correspondence problem, i.e. matching and tracking. These topis will be attack by means of statistic filtering, invariant theory. In the end, we expect the gained theoretical insights to induce novel laws for the control of robots by visual feedback with improved performance and flexibilty.
Training content: Formal objective: PhD-degree from the INPG, Grenoble. Content:
Acquire theoretical knowledge in projective geometry and control theory. (through teaching and research) Complement knowledge in computer vision. (through research) Specialize, in-field experience in visual servoing. (through research and practice)
Links with industry: Industrial colaboration: (through LTR-Esprit VIGOR industrial partners) exchange experience, workshops in robotics share, evaluate sample test data provide testbeds, dissemination of results