DiT-SLAM: Real-Time Dense Visual-Inertial SLAM with Implicit Depth Representation and Tightly-Coupled Graph...

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Real-time dense map generation is crucial for mobile robotics.
• Existing visual-inertial SLAM systems have limitations in local time windows or loosely-coupled frameworks.
• Implicit depth representations show promise but require better integration into SLAM.

Purpose of the Study:

• To develop a novel real-time dense visual-inertial SLAM system.
• To address limitations of filter-based and loosely-coupled SLAM frameworks.
• To improve dense map generation and state estimation accuracy.

Main Methods:

• Proposed DiT-SLAM: Dense visual-inertial SLAM with implicit depth representation and tightly-coupled graph optimization.
• Developed a light-weight monocular depth estimation and completion network with attention and CVAE.
• Introduced a robust point sampling strategy for geometric constraints in challenging environments.

Main Results:

• Simultaneous optimization of poses, sparse maps, and depth codes using visual, inertial, and depth residuals.
• Uncertainty-aware dense depth maps generated from low-dimensional codes.
• Improved performance in both dense depth estimation and trajectory estimation compared to baselines.

Conclusions:

• DiT-SLAM offers a robust and accurate solution for real-time dense visual-inertial SLAM.
• The tightly-coupled graph optimization framework effectively integrates multi-modal sensory data.
• The proposed depth estimation network and sampling strategy enhance SLAM performance, especially in feature-poor conditions.