Probabilistic Depth Map Model for Rotation-Only Camera Motion in Semi-Dense Monocular SLAM

Handing rotation-only camera motion is challenge work for current direct (featureless) monocular SLAM with six degrees of freedom. In rotation-only camera motion, existing systems can't estimate and update depth map and will get tracking failure. In this paper, a probabilistic depth map model is proposed based on Bayesian estimation in a per-pixel level. Both rotation-only and general camera motion can be tracked and mapped with this depth map model in real-time. The experiment results illustrate that our approach can handles both general and rotation-only camera motion and creates a larger and denser maps compared to the state-of-art semi-dense SLAM system.