The development of the Autonomous Guided Vehicles (AGVs) with urban applications are now possible due to the recent solutions (DARPA Grand Challenge) developed to solve the Simultaneous Localization And Mapping (SLAM) problem: perception, path planning and control. For the last decade, the introduction of GPS systems and vision have been allowed the transposition of SLAM methods dedicated to indoor environments to outdoor ones. When the GPS data are unavailable, the current position of the mobile robot can be estimated by the fusion of data from odometer and/or Inertial Navigation System (INS). We detail in this article what can be done with an uncalibrated stereo-rig, when it is embedded in a vehicle which is going through urban roads. The methodology is based on features extracted on planes: we mainly assume the road at the foreground as the plane common to all the urban scenes but other planes like vertical frontages of buildings can be used if the features extracted on the road are not enough relevant. The relative motions of the coplanar features tracked with both cameras allow us to stimate the vehicle ego-motion with a high precision. Futhermore, the features which don't check the relative motion of the considered plane can be assumed as obstacles.