RDNV-SLAM: A Real-Time Dynamic NeRF-Based Vision SLAM System

In recent years, Neural Radiation Fields (NeRF)based SLAM systems have achieved good positioning accuracy and mapping results in static environments. However, in dynamic scenes, NeRF-based SLAM faces significant challenges. The pose estimation tends to drift significantly over time and fails to meet the real-time requirements of SLAM systems. To address these problems, we propose a real-time dynamic visual SLAM system based on NeRF, called RDNV-SLAM. We generate semantic masks using semantic and depth information, remove dynamic region feature points by combining optical flow with hypothesis testing, and ensure the motion consistency of the remaining features using depth-RANSAC. During the rendering and mapping process, we use the semantic masks as prior information and apply multi-resolution hash encoding to reduce the impact of dynamic objects on mapping and improve rendering efficiency. We collected and created the NPU-CameraD dataset and conducted experiments on the public datasets TUM and Bonn, as well as our own collected dataset. Compared with existing classical SLAM algorithms, our method achieves better tracking accuracy and mapping quality in dynamic scenes and better meets the real-time requirements.