Merging Partially Consistent Maps

Learning maps from sensor data has been addressed since more than two decades by Simultaneous Localization and Mapping (SLAM) systems. Modern state-of-the-art SLAM approaches exhibit excellent performances and are able to cope with environments having the scale of a city. Usually these methods are entailed for on-line operation, requiring the data to be acquired in a single run, which is not always easy to obtain. To gather a single consistent map of a large environment we therefore integrate data acquired in multiple runs. A possible solution to this problem consists in merging different submaps. The literature proposes several approaches for map merging, however very few of them are able to operate with local maps affected by inconsistencies. These methods seek to find the global arrangement of a set of rigid bodies, that maximizes some overlapping criterion. In this paper, we present an off-line technique for merging maps affected by residual errors into a single consistent global map. Our method can be applied in combination with existing map merging approaches, since it requires an initial guess to operate. However, once this initial guess is provided, our method is able to substantially lessen the residual error in the final map. We validated our approach on both real world and simulated datasets to refine solutions of traditional map merging approaches.