Representation etendue de scenes combinant des primitives intrinseques geometriques et photometriques

The aim of this work is to detect objects' surfaces in an structured environment, such as the interior of a building, and to create a virtual reconstruction of this environment which respects the positions, orientations and colors of the observed surfaces.; We use the following restrictive assumptions. The observed environment is static and most of the observed surfaces are planar. The environment's light is uniform and constant throughout the observations. Moreover, we suppose that the sensor is mounted on a platform moving on a planar surface.; We designed a new 3D sensor based on a laser rangefinder, a camera and a tilt platform in order to carry out registered telemetric and photometric data acquisition. For each acquisition, this sensor provides a grid of 61,009 positions and colors corresponding to points observed in the environment facing the platform. The obtained results show that the average telemetric error of our sensor is 5.9 mm and that the maximum shift observed between telemetric and photometric data is 1 pixel.; The reconstructed environment is made of many kind of geometric primitives which are portions of planes, spheres and cylinders. Our representation can be expanded to add other types of premitives as well. To extract these premitives, two segmentations are performed, one based on telemetry and the other based on photometry. These two segmentations are then merged to obtain a new segmentation which takes into account both sensory informations. We also show that photometry can be used in certain situations to refine the edges of telemetric segmentation. The obtained results show that our segmentation method detects correctly more than 90% of surfaces found in the environment.; Surfaces obtained after the segmentation process are integrated into a map. In order to do so; we designed a new localization algorithm to locate the sensor in its environment. This algorithm is a modified version of the ICP algorithm (Iterative Closest Points) which is often used in mobile robotics. Our modifications make it possible to use the detected surfaces to carry out the localization rather than use the raw data directly. Hence we get a considerable gain of performance compared to traditional ICP. With the integration of surfaces in a map we obtain an extended representation of the environment in which the error of the surfaces' dimension is approximately 5 cm and the error of orientation is around 0.25°.; The choice of the best viewpoints and the navigation from one viewpoint to another are carried out by an operator. Upon the operator's request, the mobile platform on whom is mounted the sensor carries out autonomously data acquisition, data segmentation, sensor localization based on the observations and integration of new surfaces to the map being built. The system we designed could be used as building block for future developments. These developments may consist of the optimized choice of viewpoints and the autonomous navigation, with obstacle avoidance, from one viewpoint to the other.