EM-SLAM With Inertial/Visual Applications

The general simultaneous localization and mapping (SLAM) problem aims at estimating the state of a moving platform simultaneously with map building of the local environment. Current state-of-the-art methods such as given by Williams et al. relies on nonlinear least-squares (NLS) batch formulations with structure exploitation for memory efficiency and speed. We investigate the expectation-maximization (EM) algorithm for solving a generalized version of the NLS problem. This EM-SLAM algorithm solves two simpler problems iteratively yielding a low computational complexity. The iterations switch between state estimation, which can use any state-space smoother, and map estimation, where a quasi-Newton method is suggested. The proposed method is evaluated in real experiments and also in simulations on a platform with a monocular camera attached to an inertial measurement unit. The results show that EM-SLAM has much lower computational complexity than NLS while maintaining comparable accuracy.