Visual-inertial motion priors for robust monocular SLAM

Abstract

Monocular visual SLAM approaches are mostly constrained in their performance due to general motion model and availability of true scale information. We proposed an approach which improves the motion prediction step of visual SLAM and results in better estimation of map scale. The approach utilizes the short term accuracy of inertial velocity with visual orientation to estimate refined motion priors. These motion priors are fused with sparse number of 3D map features to constraint the positional drift of moving platform. Experimental results are presented on large scale outdoor environment, yielding robust performance and better observability of map scale by monocular SLAM.