Handling pure camera rotation in semi-dense monocular SLAM

In this paper, we present a method for semi-dense monocular simultaneous localization and mapping (SLAM) that is capable of dealing with pure camera rotation motion which brings forward a severe challenge for current direct (featureless) monocular SLAM approaches. A probabilistic depth map model built on Bayesian estimation is combined with the main framework of the state-of-the-art direct method LSD-SLAM. Using this model, both rotation-only and general camera motions could be tracked, and a consistent depth map could be built in real-time. Experimental results demonstrate the outstanding performance of the proposed system.