The starting point of this study is the development of a robot assistant for the disabled. The robot is a vision-based controlled manipulator which is equipped with two cameras : one is embedded on the gripper and gives a close view of the scene while the second one is remotely located and gives a global view of the scene. The objective is then to grasp an a priori unknown object given the only information of one clic on the remote image. We present methods to roughly localize the object and estimate the characteristics needed for grasping. Our work may be seen as an alternative to the grasping procedures that are using a previously built data-base. The thesis is divided in two parts : the rough object localization and the estimation of its characteristics. Given the coordinates of the clicked point, the object is known to be on the view line which connects both the remote camera optical center and the clicked point. The projection of this view line on the gripper image plane is the epipolar line associated with the clicked point. Epipolar based visual servoing is used to control the embedded camera to scan this line. Image characteristics are extracted from both the remote and the gripper view and then matched to estimate the 3D position of the object. This method holds the advantage of being robust to object motion in the remote frame. At the end of the localization process, the object is included in both ﰜelds of view and the estimation of the characteristics is initialized. The object rough shape estimation is treated with a monocular mobile camera. The object shape is approximated by a quadric whose parameters are estimated from the object projection on a set of images. The object is segmented using an active contour method that is initialized using the output of the localization process. The better the viewpoints, the more accurate the characteristics estimation. Finally, a active vision method is developed to automatically select the viewpoints that improve reconstruction. The best views are chosen in order to maximize the new information.
