Loop Closure with 3D Gaussian Splatting for Dynamic SLAM.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Visual Simultaneous Localization and Mapping (SLAM) faces challenges in dynamic environments, particularly with 3D Gaussian Splatting (3DGS) methods.
• Existing 3DGS techniques struggle with static scene assumptions and inconsistencies in dynamic settings, degrading accuracy.

Purpose of the Study:

• To present a novel dynamic SLAM method, Loop Closure with 3D Gaussian Splatting for Dynamic SLAM (LCD-Splat).
• To enhance pose estimation and high-fidelity scene reconstruction in dynamic environments.

Main Methods:

• LCD-Splat integrates Mask R-CNN and multi-view geometry for dynamic object detection and static map generation.
• It utilizes 3DGS submaps and a frame-to-model tracking strategy for dense map construction.
• Online loop closure detection and a coarse-to-fine 3DGS registration algorithm are employed for global consistency via pose graph optimization.

Main Results:

• LCD-Splat demonstrates superior performance over state-of-the-art SLAM methods on real-world datasets (TUM RGB-D, Bonn).
• The method achieves improved tracking, scene reconstruction, and rendering performance in dynamic environments.

Conclusions:

• LCD-Splat offers a robust solution for high-precision SLAM in dynamic settings.
• The approach provides valuable insights for scene understanding in complex, real-world scenarios.