Exactly Rao-Blackwellized unscented particle filters for SLAM

This paper addresses the limitation of the conventional Rao-Blackwellized unscented particle filters. The problem is on the usage of the overconfident optimal proposal distribution caused by perfect map assumption, so that predictive robot poses are sampled from the underestimated error covariance in the particle filtering process. The proposed solution computes more precise error covariance of the robot which contains uncertainties of the robot, map, and measurement noise. Experimental results using the benchmark dataset confirmed that the covariance of the proposed method is always larger than that of the conventional method while inducing slower increasing rate of the weight variance with less resamplings.