Online Expectation Maximization algorithm to solve the SLAM problem

摘要

In this paper, a new algorithm - namely the onlineEM-SLAM - is proposed to solve the simultaneous localization and mapping problem (SLAM). The mapping problem is seen as an instance of inference in latent models, and the localization part is dealt with a particle approximation method. This new technique relies on an online version of the Expectation Maximization (EM) algorithm: the algorithm includes a stochastic approximation version of the E-step to incorporate the information brought by the newly available observation. By linearizing the observation model, the stochastic approximation part is reduced to the computation of the expectation of additive functionals of the robot pose. Therefore, each iteration of the onlineEM-SLAM both provides a particle approximation of the distribution of the pose, and a point estimate of the map. This online variant of EM does not require the whole data set to be available at each iteration. The performance of this algorithm is illustrated through simulations using sampled observations and experimental data.