MBA-SLAM: Motion Blur Aware Dense Visual SLAM With Radiance Fields Representation.

Area of Science:

• Computer Vision
• Robotics
• 3D Scene Reconstruction

Background:

• Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) excel in SLAM with high-quality video.
• Motion-blurred frames, common in real-world scenarios, degrade SLAM performance, impacting localization and reconstruction.
• Existing methods lack robustness when dealing with severe motion blur.

Purpose of the Study:

• To develop a dense visual SLAM pipeline capable of handling severe motion-blurred inputs.
• To improve camera localization accuracy and 3D map reconstruction quality under motion blur.
• To enable robust SLAM in challenging real-world conditions like low-light or long-exposure photography.

Main Methods:

• Proposed MBA-SLAM, a motion blur-aware SLAM pipeline integrating an efficient tracker with NeRF or 3DGS mappers.
• Modeled the physical image formation process of motion-blurred images.
• Simultaneously learned 3D scene representation and estimated camera trajectory during exposure time for motion blur compensation.

Main Results:

• MBA-SLAM demonstrated superior performance over state-of-the-art methods in both camera localization and map reconstruction.
• Achieved significant improvements on datasets with motion-blurred images, outperforming existing techniques.
• Showcased robustness and versatility across synthetic and real-world datasets, including those with sharp and blurred images.

Conclusions:

• MBA-SLAM effectively addresses the challenge of motion blur in visual SLAM.
• The proposed method offers a robust solution for accurate 3D scene representation and camera localization in real-world scenarios.
• MBA-SLAM significantly advances the capabilities of NeRF and 3DGS-based SLAM systems in the presence of motion blur.