The combination of SfM and monocular SLAM

To realize autonomous navigation with a regular camera in an unknown environment, there are mainly two types of approaches: SfM and monocular SLAM. They both have advantages and disadvantages. SfM is slow and cannot eliminate outliers, but it is able to provide 3D information from a series of images without any additional information. Monocular SLAM can hardly work unless the initial value is close to the real value, but it is fast and can handle outliers naturally. The combination approach proposed in this paper combines SfM and monocular SLAM. It uses SfM as a observer to linearize the observation function used in monocular SLAM, and uses the results of monocular SLAM to accelerate SfM. Outliers are pointed out by SfM and handled by monocular SLAM. Simulation and experiment results show that the proposed combination approach is feasible. The accuracy is improved compared with SfM and outliers can be eliminated.