A robust visual SLAM method based on point feature fusion

In visual SLAM (simultaneous localization and mapping) based on conventional point features, factors like rapid perspective changes of camera and environment with weak textures and unclear features can cause the system to fail or even make it impossible to initialize. To solve this problem, we propose a SLAM system with multiple feature points fusion. First, in this system, ORB and GCNv2 features are used in combination to ensure the robustness of tracking threads. Secondly, a new condition for generating keyframes is proposed, which uses relative rotation and translation between frames to determine whether keyframes should be created. Finally, the pose propagation between maps is carried out before the global BA optimization to improve the optimization speed and the reliability of mapping, and make the final map generated by the SLAM system have global consistency. We used the framework of ORB-SLAM3 as a carrier and conducted experiments on Tum datasets and real environments. The experimental results show that the proposed algorithm can complement the advantages of various feature points on the premise of ensuring good accuracy, and effectively improve the robustness of the SLAM system in environments with weak textures, unclear features, and camera’s perspective changes rapidly.