This document concerns service robotics for human assistance. A companion robot will have to manipulate everyday 3D objects (bottles, glasses...), recognized and localized from data acquired with sensors embedded on the robot, here using monocular or stereo vision. For vision-based object manipulation, it is necessary first to learn two representations for every object; a 3D geometrical model, mandatory to control the grasping task, and an appearance-based model, required for the visual recognition. This thesis deals first with the construction of these representations, and then proposes an active method for object recognition from images acquired from embedded cameras. The modeling is performed on a 3D object set alone on a table; 3D data are acquired from a stereo rig mounted on a manipulator; the sensor is moved by the arm around the object in order to acquire N images, from which a triangular mesh is built. It is proposed first an original approach for the registration of partial views, approach based on a pseudo-color created from the 3D points acquired on the object surface. Then an efficient method, based on a spherical parametrization, is proposed to make simpler the construction of a triangular mesh from the registered views aggregated in a 3D points cloud. The active recognition method is based on a single camera. The learning of the appearance-based model is also built, moving the camera around every object set alone on a table. This model is made of several views: for everyone, (1) the object silhouette is first extracted using a snake, (2) then, several descriptors are computed, either global (color, silhouette signature, shape context computed on all the object region) or local ones (interest points, color or shape contexts in discretized regions). The recognition process analyzes a scene with a single object, or with several ones set without order, including unknown objects. An incremental active method allows to update a probability vector P(Obji), i=1, N+1 if N objects have been learnt; the unknown objects are assigned to the class N+1; P(Obji) gives the probability that an object from the class i is in the scene. After every step, the best view point is selected for the next sensor position, using the maximization of the mutual information. The method has been validated from numerous results from synthetic or true images.
