RBPF-SLAM based on probabilistic geometric planar constraints

Recent advances in visual SLAM have focused on improving estimation of sparse 3D points or patches that represent parts of surroundings. In order to establish an adequate scene understanding, inference of spatial relations among landmarks must be part of the SLAM processing. A novel Rao-Blackwilized PF-SLAM algorithm is proposed to utilize geometric relations of landmarks with respect to high level features, such as planes, for improving estimation. These geometric relations are defined as a set of geometric constraint hypotheses inferred during the mapping task. In each prediction-update cycle of estimation, probabilistic constraints are created and applied to update landmarks based on a hierarchical inference process. Based on experiments, improvement over estimation and completeness of the scene description is achieved using the proposal of this paper.