Inferring and Enforcing Relative Constraints in SLAM

Most algorithms for simultaneous localization and mapping (slam) do not incorporate prior knowledge of structural or geometrical characteristics of the environment. In some cases, such information is readily available and making some assumptions is reasonable. For example, it is often safe to assume that many walls in an indoor environment are rectilinear. In this paper, we develop a slam algorithm that incorporates prior knowledge of relative constraints between landmarks. We describe a “Rao-Blackwellized constraint filter” that infers applicable constraints and efficiently enforces them in a particle filtering framework. We have implemented our approach with rectilinearity constraints. Results from simulated and real-world experiments show the use of constraints leads to consistency improvements and a reduction in the number of particles needed to build maps.