Point set registration based on multi-object metrics

Abstract

In robotics and computer vision, point set registration is necessary in many tasks, for example in estimating the motion of a sensor/sensors between subsequent scans containing point/feature sets. Currently, the Iterated Closest Point (ICP) method and its variants have been presented as possible solutions to this problem. However most of these methods lack robustness when random spatial and detection errors are present. This is because ICP methods typically use an L2 metric as part of their optimization criteria, which is unable to penalize cardinality errors. Therefore, this article presents a registration technique based on the multi-object Optimal Sub-Pattern Assignment (OSPA) and Cardinalized Optimal Linear Assignment (COLA) metrics, which penalize data differences based on both cardinality and spatial errors. This allows scan registration to take place in the presence of both inter-scan translation and orientation as well as detection errors.