SVR-Net: A Sparse Voxelized Recurrent Network for Robust Monocular SLAM with Direct TSDF Mapping

Area of Science:

• Robotics and Computer Vision
• Artificial Intelligence and Machine Learning

Background:

• Simultaneous Localization and Mapping (SLAM) is crucial for robot navigation and planning.
• Monocular visual SLAM systems struggle with robust pose estimation and map generation.

Purpose of the Study:

• To propose SVR-Net, a robust monocular SLAM system utilizing a sparse voxelized recurrent network.
• To improve pose estimation accuracy and enable efficient dense map construction for downstream tasks.

Main Methods:

• SVR-Net employs a sparse voxelized recurrent network for feature extraction and correlation.
• Gated recurrent units iteratively refine matches on correlation maps for enhanced robustness.
• Gauss-Newton updates are integrated to enforce geometrical constraints for precise pose estimation.

Main Results:

• SVR-Net successfully estimated poses across all nine scenes in the TUM-RGBD dataset.
• Performance demonstrated comparable tracking accuracy to DeepV2D, outperforming traditional ORB-SLAM.
• The system directly generated dense TSDF maps, efficient for subsequent applications.

Conclusions:

• SVR-Net offers a robust and accurate solution for monocular visual SLAM.
• The system's sparse voxelization and recurrent architecture improve efficiency and performance.
• Direct dense TSDF map generation enhances usability for navigation and planning tasks.