Automatic Relocalisation for a Single-Camera Simultaneous Localisation and Mapping System

We describe a fast method to relocalise a monocular visual SLAM (simultaneous localisation and mapping) system after tracking failure. The monocular SLAM system stores the 3D locations of visual landmarks, together with a local image patch. When the system becomes lost, candidate matches are obtained using correlation, then the pose of the camera is solved via an efficient implementation of RANSAC using a three-point-pose algorithm. We demonstrate the usefulness of this method within visual SLAM: (i) we show tracking can reliably resume after tracking failure due to occlusions, motion blur or unmodelled rapid motions; (ii) we show how the method can be used as an adjunct for a proposal distribution in a particle filter framework; (iii) during successful tracking we use idle cycles to test if the current map overlaps with a previously-built map, and we provide a solution to aligning the two maps by splicing the camera trajectories in a consistent and optimal way.