ORB-SLAM map initialization improvement using depth

Map initialization and scale ambiguity are well-known challenging problems for Visual SLAM. ORB-SLAM authors [1] [2] [3] [4] have published automatic initialization and excellent Bundle Adjustment. However, scale ambiguity remains an open issue. Since depth sensors provide scale information, RGB-D SLAM improves this scale ambiguity problem. However, RGB-D SLAM depth estimation is provided only within limited areas and the measurement error usually increases with the distance. To avoid this problem, in this paper a triangulation is used on RGB feature points for getting 3D points from out of the limited area in depth. We combined both advantages of triangulation and depth. Our RGB-D ORB-SLAM system is evaluated in TUM RGB-D dataset [5] and compared with ORB-SLAM. Our analysis concludes to a better robustness to initialization and tracking.