Comparing the EKF and FastSLAM solutions to the problem of monocular Simultaneous Localization and Mapping.

Abstract

This paper presents a comparison of the extended Kalman filter (EKF-SLAM) and FastSLAM algorithms, the two most popular solutions to the simultaneous localization and mapping problem (SLAM). We focus strictly on the class of monocular indoor mobile robots. Because the extended Kalman filter only maintains one solution, it is known to be very fragile to incorrect data-associations. FastSLAM, on the other hand, has it’s background in par-ticle filtering, in which it tracks multiple hypotheses at the same time, which enable it do be robust under incorrect data-associations. Most papers comparing the two algo-rithms do this by means of simulation, or on real-world data with poor ground truth, making it difficult to esti-mate the performance of the filter. We however compare both algorithms on real-world data with high-precision ground truth available. The experiments show that the extended Kalman filter suffers significantly from incorrect data-associations. FastSLAM does not suffer significantly from the incorrect data-associations and is thus a robust filter in situations where the data-association problem is very hard. We also show that, despite the better perfor-mance, FastSLAM is less consistent compared to EKF-SLAM. 1