In this thesis, we develop visual SLAM system that combines classical computer vision principles with state-of-the-art deep learning approaches. The key feature of our implementation is the use of RoMa, a dense feature matcher based on neural networks, enabling more robust and precise tracking of camera motion through space. Additionally, we perform local and global trajectory optimization and loop closure using the AnyLoc neural model, further improving the accuracy and consistency of the generated map. The system is evaluated on standard benchmark dataset TUM RGB-D, allowing for quality comparison with existing approaches. The analysis of the results showed that the designed hybrid architecture operates reliably and achieves competitive accuracy, which confirms the potential of such approaches and opens new directions for future research.
|
