Hierarchical SLAM using spectral submap matching with opportunities for long-term operation

We present a hierarchical SLAM approach which uses spectral registration of local submaps to close loops and to perform global localization after a restart. Using the Fourier-Mellin Transform (FMT), we robustly register occupancy grid representations of local submaps and present methods which improve matching performance. We further show how good match candidates can be reliably detected even from scaled-down versions of the submaps, which significantly reduces the computation time. The spectral registration approach proves useful even in the presence of significant environmental changes due to the fact that it calculates a dense match, incorporating all observed information rather than a sparse set of features.