Efficient large scale SLAM including data association using the Combined Filter

摘要

In this paper we describe the Combined Filter, arnjudicious combination of Extended Kalman (EKF) and ExtendedrnInformation filters (EIF) that can be used to execute highlyrnefficient SLAM in large environments. With the CF, filterrnupdates can be executed in as low as O(log n) as comparedrnwith other EKF and EIF based algorithms: O(n~2) for MaprnJoining SLAM, O(n) for Divide and Conquer (D&C) SLAM,rnand O(n~(1.5)) for the Sparse Local Submap Joining Filter (SLSJF).rnWe also study an often overlooked problem in computationallyrnefficient SLAM algorithms: data association. In situations inrnwhich only uncertain geometrical information is available forrndata association, the CF Filter is as efficient as D&C SLAM,rnand much more efficient than Map Joining SLAM or SLSJF.rnIf alternative information is available for data association, suchrnas texture in visual SLAM, the CF Filter outperforms all otherrnalgorithms. In large scale situations, both algorithms based onrnExtended Information filters, CF and SLSJF, avoid computingrnthe full covariance matrix and thus require less memory, butrnstill the CF Filter is the more computationally efficient. Bothrnsimulations and experiments with the Victoria Park dataset, thernDLR dataset, and an experiment using visual stereo are used tornillustrate the algorithms’ advantages.