Existing methods based on local features cannot recognize objects in real-time while keeping a high recognition rate.Considering that many local features are unstable，unreliable，or irrelevant，we are able to select a small subset of features used for recognition by correctly matching features in training images.A new，robust，and stable method based on a bag-of-features is proposed in this paper.Distinctive features are selected by an unsupervised preprocessing step.Our experiments demonstrate that this selection approach can reduce the amount of local features and reduce the memory requirements，while allowing an average of 4%of the original features per image to provide matching performance that is as accurate as the full set.The method can meet real-time requirements since the time required for matching has been reduced from seconds to tens of milliseconds.