SD-VIS: A Fast and Accurate Semi-Direct Monocular Visual-Inertial Simultaneous Localization and Mapping (SLAM)

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Balancing accuracy and computational efficiency is a key challenge in Simultaneous Localization and Mapping (SLAM).
• Existing SLAM systems often struggle to achieve both high performance and real-time processing.

Purpose of the Study:

• To propose SD-VIS, a novel fast and accurate semi-direct visual-inertial SLAM framework.
• To enhance camera motion estimation and sparse scene structure reconstruction.
• To achieve a better trade-off between accuracy and computational efficiency in SLAM.

Main Methods:

• A hybrid front-end combining feature-based tracking for keyframes and direct method for non-keyframes.
• A sliding window-based, tightly-coupled optimization back-end with marginalization for computational complexity control.
• Initialization using multi-view geometry (MVG) for scale, velocity, gravity, and bias calibration.

Main Results:

• SD-VIS demonstrates superior real-time performance on the EuRoC dataset.
• The framework achieves a more accurate state estimation by minimizing visual and IMU errors.
• Experimental evaluation confirms the feasibility and efficiency of the proposed system.

Conclusions:

• SD-VIS offers a significant improvement in balancing accuracy and speed for visual-inertial SLAM.
• The proposed semi-direct approach effectively leverages the strengths of both feature-based and direct methods.
• SD-VIS presents a viable solution for real-time applications requiring robust localization and mapping.