Monocular SLAM Based on Semi-Direct method

We propose a monocular SLAM based on semi-direct method, which combines the advantages of direct method and indirect (feature-based) method. On the basis of extracting the salient feature points of the image, the pose of the camera is estimated by the direct method. The key-frame descriptors are represented by image description model based on learning. The key-frame descriptors are combined with the effective retrieval strategy to get rough closed-loop detection, and the descriptors are combined with direct matching to verify closed-loop candidate. Compared with popular open source algorithms, our method is tested on the publicly available datasets. Compared with state-of-the-art direct-based method and indirect-based method slam, our experimental results improve the overall performance of the system.