A Robust Framework Fusing Visual SLAM and 3D Gaussian Splatting with a Coarse-Fine Method for Dynamic Region...

Area of Science:

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

Background:

• Existing visual SLAM struggles with dynamic environments due to moving objects.
• Neural representations perform well in static scenes but degrade with motion.

Purpose of the Study:

• To develop a robust dynamic SLAM framework for accurate localization and dense mapping in environments with moving objects.
• To improve the performance of SLAM systems in challenging dynamic scenarios.

Main Methods:

• Combines classic geometric features for localization with learned photometric features for dense mapping.
• Utilizes instance segmentation and Kalman filters for object tracking.
• Employs a cascaded, coarse-to-fine optical flow strategy for efficient motion analysis.
• Filters features in dynamic regions to enhance camera pose estimation.
• Uses a 3D Gaussian Splatting backend with a Gaussian pyramid for high-quality reconstruction.

Main Results:

• Achieved up to 95% reduction in Absolute Trajectory Error on dynamic datasets compared to ORB-SLAM3.
• Generated clean and high-fidelity dense maps in dynamic scenarios.
• Reduced motion analysis computation by 91.7% compared to dense-only methods.
• Demonstrated robustness and accuracy across diverse datasets.

Conclusions:

• The proposed dynamic SLAM framework significantly enhances localization and mapping accuracy in dynamic environments.
• The system offers a robust and efficient solution for real-world applications involving moving objects.
• The integration of geometric and learned features provides a powerful approach for dynamic SLAM.