Neural Radiance Field Dynamic Scene SLAM Based on Ray Segmentation and Bundle Adjustment

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Neural implicit SLAM excels in static scenes but struggles with dynamic environments and lighting variations.
• Handling real-world scenes with motion and illumination changes is a key challenge in SLAM.

Purpose of the Study:

• To develop a neural implicit SLAM method capable of robustly handling dynamic scenes.
• To improve the accuracy and quality of 3D reconstructions in challenging, real-world scenarios.

Main Methods:

• A keyframe selection and tracking switching approach using Lucas-Kanade (LK) optical flow.
• A semantic-based joint estimation for dynamic and static pixels using Conditional Random Fields (CRFs).
• Development of constrained loss functions for dynamic scene modeling.

Main Results:

• The proposed method demonstrates superior performance on dynamic scene datasets (TUM RGB-D, Openloris, Bonn).
• Significant improvements in reconstruction quality compared to existing neural implicit SLAM systems.
• Enhanced tracking accuracy in challenging, dynamic environments.

Conclusions:

• The novel neural implicit SLAM approach effectively addresses limitations in dynamic scene reconstruction.
• The method offers a robust solution for real-world SLAM applications with motion and lighting variability.
• This work advances the state-of-the-art in neural implicit SLAM for complex environments.