The Scale Invariant Feature Transform, SIFT, has been successfully applied to robot vision, object recognition, motion estimation, etc. In this work, we propose a SIFT improvement that makes feature extraction and matching more robust, adding a feature group matching layer, which takes into account mutual spatial relations between features. The feature group matching is very fast to be computed and leads to interesting results, above all for the absence of outliers. Results of vision based robot localization using the proposed approach are presented.