In appearance-based localization, the robot environment is implicitly represented as a database of features derived from a set of images collected at known positions in a training phase. For localization the features of the image, observed by the robot, are compared with the features stored in the database. In this paper we propose the application of the integral invariants to the robot localization problem on a local basis. First, our approach detects a set of interest points in the image using a Difference of Gaussian (DoG)-based interest point detector. Then, it finds a set of local features based on the integral invariants around each of the interest points. These features are invariant to similarity transformation (translation, rotation, and scale). Our approach proves to lead to significant localization rates and outperforms a previous work.