Map matching compiles small numbers of consecutive scans into local maps \(m_{\mathrm{local}}\). Maps are then compared to the global map \(m\), such that the more similar \(m\) and \(m_{\mathrm{local}}\), the larger \(p(m_\mathrm{local} | x_t, m)\). Cells in the local map are transformed to the coordinates of the global map, and the map correlation function is computed.

Map matching is easy to compute, but does not yield smooth probabilities in the pose parameter \(x_t\). One solution is to convolve the map with a Gaussian smoothness kernel before running map matching.

Map matching considers the free-space in the scoring of 2 maps, compared to the §likelihood_field_model which uses only the end-points.

## Issues

- No plausible physical explanation
- Result of map matching may incorporate areas beyond actual measurement range.