Contributions a la modelisation et a la commande en asservissement visuel

摘要

Our work is in the field of robotic vision, and more precisely that of visual servoing. It consists of designing methods which require neither knowledge of the model of the observed object nor precise calibration of the system. We first studied the problem of Euclidean reconstruction (up to a scale factor) from two images of a static object whose model is unknown. We have shown that the best linear solution to obtain a robust reconstruction of the camera movement is to use a homography matrix. An algorithm for the linear estimation of this was then proposed. Then, we proposed two new visual servoing methods based on partial Euclidean reconstruction. We studied the closed-loop system in order to analytically determine the robustness of the control laws to uncertainties in the calibration of the system. Then, we showed that, thanks to the decoupled form of the systems considered, it is possible to significantly increase the robustness domain with an adaptation of the control gains. The new methods were tested on a Cartesian robot and a redundant robot, and compared to traditional methods. The results obtained confirm that our control schemes provide a significant improvement in the convergence domain. Finally, we extended the servo schemes to the use of several cameras not observing the same part of the scene. We modeled multi-camera visual servoing within the framework of sensor-referenced control. The particular choice of the spot function allowed us to simplify the control design and stability analysis. We finally carried out a task of positioning a bulky object using two cameras on board a manipulator robot.