This paper describes a technique for image stabilization in video sequences. The warping that compensates for camera's motion is computed from tracked features in the images. In order to cope with moving objects, a robust technique is used to compute homographies. Moreover, the tracking is made more reliable by using the computed warping to help predicting the features' positions. The effectiveness of the motion compensation is demonstrated by constructing mosaic images from the stabilized sequence and by computing the RMS error. An effort has been made to keep the computational cost low and to reduce the frame rate needed for tracking, with the aim to make a real-time implementation viable.