An algorithm for feature point extraction is presented. It is based on a scale-space representation of the image as well as a system for tracking across scales. Using synthetic and real images, it is shown that the proposed algorithm produces stable and well-localized feature points estimates two essential properties for video applications.