Research on Monocular Vision SLAM Algorithm for Multi-map Fusion and Loop Detection

Aiming at the low robustness of feature point SLAM to scene location and mapping with missing features such as occlusion, image frame damage and low texture, a monocular visual SLAM (MMF-SLAM) algorithm based on multi-map fusion and loop detection is proposed in this paper. In this paper, the local multi-map fusion method is used to solve the problem of choosing fast relocation or constructing a new local map from the current position when the tracking is lost, and then fuse all the previously constructed maps when looping back. Secondly, in the loop detection module, the TF-IDF score matching standard is improved, a new location recognition technology with high recall is proposed, and a more detailed and accurate map fusion is carried out. The test results on the TUM RGB-D benchmark dataset and the KITTI dataset show that the multi-map fusion method effectively solves the problem that the map cannot be built after the tracking is lost. At the same time, the improved loop detection reduces the error accumulation of the system and makes the system have good positioning accuracy. Compared with the mainstream visual SLAM algorithm ORB-SLAM2, this algorithm can quickly relocate and continuously track and map fusion, which helps to improve the calculation accuracy of SLAM algorithm.