RB2 — PF : A novel filter-based monocular visual odometry algorithm

This paper proposes an improvement to FastSLAM. The approach is applicable when the dynamic model describing the motion of the camera has linear sub-structure. The core novelty of the proposed algorithm is to separate the consideration of the camera's dynamic model into two sub-models without constraining the two sub-models to have independent noise processes. In contrast to commonly-used FastSLAM algorithms, which use a particle filter to consider both these sub-models, a particle filter is used for one sub-model and a Kalman filter for the other. This tactic is Rao-Blackwellisation and is the same as that which underpins the development of FastSLAM, but where the focus was only on exploiting near-linear sub-structure related to the state of the landmarks. Comparisons with Unscented FastSLAM 2.0 indicate that the new approach improves estimation accuracy. Comparisons with cutting-edge SLAM algorithms also reflect the competitive nature of this approach as a solution to navigation problems. Future work will improve the processing of features and consider multi-sensor data input.