This paper presents a Visual EKF-SLAM process using an original and very efficient strategy for initializing landmarks. Usually, with Cartesian coordinates, new points are created along the line-of-sight with a large variance. However, this type of initialization is subject to significant linearization issues making landmarks diverge from their real position. The immediate consequence is a failure of the Visual SLAM process. We propose here a new strategy that avoids or drastically limits the linearization errors. The first part of this strategy takes place during the tracking process where a coherent window is needed in order to successfully follow a point and make it converge. The second part concerns the update step. Due to linearization errors, a landmark in front of the observer can be updated behind it. We compute a corrective of the Kalman gain in order to preserve the integrity. We applied this strategy to real data illustrating its efficiency.