Visual inertial SLAM algorithm with SuperPoint and semantic information

In order to address the challenges faced by traditional visual SLAM algorithms in complex scenarios, such as low texture and dynamic objects, this paper proposed a robust SLAM system that improves the ORB-SLAM3 framework. The algorithm used the SuperPoint network for feature extraction and introduced a dynamic confidence threshold module to obtain a robust set of feature points in challenging scenes. Additionally, it detected and classified objects within the scene by using the YOLO11-seg semantic segmentation network, and it proposed a geometric constraint method combining depth information to eliminate the feature points of dynamic objects. Furthermore, it leveraged LightGlue for high-confidence feature matching and employed deep local feature descriptors to replace traditional handcrafted features, enabling fast and accurate matching. Finally, the algorithm is validated on the TUM and EuRoC datasets. The results show that compared to traditional visual SLAM algorithms like ORB-SLAM3 and other deep learning-related SLAM algorithms, the proposed algorithm has better positioning accuracy in complex and challenging scenarios.