In this paper, we present our work to try and implement a SLAM algorithm on a humanoid robot platform, the NAO robot produced by Aldebaran Robotics. We first start by testing a visual SLAM algorithm which uses keypoints as visual landmarks and tries to estimate their positions, and adapt it to the specific constraints of the platform: restricted CPU, monocular camera, low speed and drifting odometry. We conclude that running a full monocular visual SLAM on the robot is not yet available, but that some specific keypoints can be robustly tracked even while walking. Then we use them to derive the robot orientation and build a compass feature based on the robot camera, which can be used for example to ensure that the robot walks straight. Therefore we showed vision can be efficiently used to improve NAO's navigation.