This dissertation presents a vision-based localization system for a mobile robot in an urban context. In this goal, the robot is first manually driven to record a learning image sequence. These images are then processed in an off-line way to build a 3D map of the area. Then vehicle can be —either automatically or manually— driven in the area and images seen by the camera are used to compute the position in the map. In contrast to previous works, the trajectory can be different from the learning sequence. The algorithm is indeed able to keep localization in spite of important viewpoint changes from the learning images. To do that, the features are modeled as locally planar features —named patches— whose orientation is known. While the vehicle is moving, its position is predicted and patches are warped to model the viewpoint change. In this way, matching the patches with points in the image is eased because their appearances are almost the same. After the matching, 3D positions of the patches associated with 2D points on the image are used to compute robot position. The warp of the patch is computationally expensive. To achieve real-time performance, the algorithm has been implemented on GPU architecture and many improvements have been done using tools provided by the GPU. In order to have a pose prediction as precise as possible, a motion model of the robot has been developed. This model uses, in addition to the vision-based localization, information acquired from odometric sensors. Experiments using this prediction model show that the system is more robust especially in case of image loss. Finally many experiments in real situations are described in the end of this dissertation. A differential GPS is used to evaluate the localization result of the algorithm.
