An incremental and absolute mobile robot self localization method +in a partially modelled indoor environment is presented. A wire frame representation of the environment is adopted. The notion of occlusion is taken into account using View Invariant Regions. A pin-hole model of the camera is obtained thanks to Zhang calibration method. The localization approach is composed of four steps : image acquisition frome robot current position and orientation, image feature extraction, 3-D/2-D feature matching and camera pose recovery. Two full perspective camera pose recovery methods using straight line correspondances and numerical optimisation technic are presented. Adaptation of these methods to mobile robotics context is defined. Finally, the crucial problem of matching image features to model features is achieved using an algorithm based on Interpretation Tree Search. Dimension of correspondance space is reduced using View Invariant Regions and the specific configuration of the robot. Two geometric constraints are used to efficiently prune the Interpretation Tree testing the local consistency. A new function is defined to test the global consistency and to select the best matching hypothesis.
