In this article, we present our work to provide a navigation and localization system on a constrained humanoid platform, the NAO robot, without modifying the robot sensors. First, we tried to implement a simple and light version of classic monocular Simultaneous Localization and Mapping algorithms, while adapting to the CPU and camera quality, which turned out to be insufficient on the platform for the moment. From our work on keypoints tracking, we identified that some keypoints can be still accurately tracked (...) at little cost, and used them to build a visual compass. This compass was then used to correct the robot walk because it makes it possible to control the robot orientation accurately. (shrink)