BDGS-SLAM: A Probabilistic 3D Gaussian Splatting Framework for Robust SLAM in Dynamic Environments.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Simultaneous Localization and Mapping (SLAM) is crucial for robotic navigation and augmented reality.
• 3D Gaussian Splatting (3DGS) offers real-time, high-fidelity rendering for SLAM but struggles with dynamic objects.
• Existing 3DGS-SLAM methods face mapping errors and tracking drift in dynamic environments.

Purpose of the Study:

• To propose BDGS-SLAM, a Bayesian Dynamic Gaussian Splatting SLAM framework for dynamic environments.
• To enhance the accuracy and robustness of SLAM in the presence of moving objects.
• To improve the fidelity of scene reconstruction and reduce rendering artifacts in dynamic scenarios.

Main Methods:

• Integrating YOLOv5 semantic detection with Bayesian filtering to create a dynamic prior probability model for identifying dynamic Gaussians.
• Employing a multi-view probabilistic update mechanism with an exponential decay factor to restore erroneously culled static Gaussians.
• Implementing an adaptive dynamic Gaussian optimization strategy to mitigate the impact of dynamic elements on rendering and preserve scene integrity.

Main Results:

• BDGS-SLAM achieves comparable tracking accuracy to baseline methods in dynamic environments.
• The proposed method generates fewer rendering artifacts compared to existing 3DGS-SLAM approaches.
• BDGS-SLAM demonstrates higher-fidelity scene reconstruction, effectively handling dynamic objects.

Conclusions:

• BDGS-SLAM provides a robust solution for SLAM in dynamic environments.
• The framework successfully addresses the limitations of traditional 3DGS-SLAM in handling moving objects.
• BDGS-SLAM offers improved scene reconstruction quality and tracking stability for real-world applications.