Monte Carlo localisation generally requires a metrical map of the environment to calculate a robots position from the posterior probability density of a set of weighted samples. Image-based localisation, which matches a robots current view of the environment with reference views, fails in environments with perceptual aliasing. The method we present in this paper is experimentally demonstrated to overcome these disadvantages in a large indoor environment by combining Monte Carlo and image-based localisation. It exploits the properties of the Fourier transform of omnidirectional images, while weighting the samples according to the similarity among images. We also introduce a novel strategy for solving the “ kidnapped robot problem”.