Direct Depth SLAM: Sparse Geometric Feature Enhanced Direct Depth SLAM System for Low-Texture Environments.

Area of Science:

• Robotics
• Computer Vision
• Simultaneous Localization and Mapping (SLAM)

Background:

• Accurate real-time pose estimation is crucial for autonomous systems.
• Depth cameras offer rich environmental data but present unique SLAM challenges.
• Existing SLAM methods often struggle with drift and robustness in real-world scenarios.

Purpose of the Study:

• To develop a real-time, robust, and low-drift depth-only SLAM system for depth cameras.
• To enhance frame-to-frame pose estimation accuracy and reduce accumulated drift.
• To achieve reliable loop closure detection and global pose optimization.

Main Methods:

• A three-layer optimization framework: Direct Depth layer for fast pose estimation using range flow, ICP Refined layer for local drift reduction, and Graph Optimization layer for global pose refinement.
• Utilizing dense range flow and sparse geometry features from sequential depth images.
• Implementing a novel loop closure detection algorithm based on sparse geometric feature matching.

Main Results:

• The proposed Direct Depth layer significantly outperforms classic methods in frame-to-frame pose estimation.
• The ICP Refined layer effectively reduces local drift through a local map-based motion estimation strategy.
• The Graph Optimization layer robustly detects loop closures and minimizes global drift, validated on benchmark and real-world datasets.

Conclusions:

• The developed depth-only SLAM method offers superior real-time performance and robustness compared to existing approaches.
• The integrated optimization layers effectively address challenges of drift and loop closure in depth camera-based SLAM.
• This work contributes a significant advancement in visual SLAM for applications relying on depth sensing.