MotionGS : Compact Gaussian Splatting SLAM by Motion Filter

Xinli Guo,Weidong Zhang,Ruonan Liu,Peng Han,Hongtian Chen

2024-05-31

Abstract:With their high-fidelity scene representation capability, the attention of SLAM field is deeply attracted by the Neural Radiation Field (NeRF) and 3D Gaussian Splatting (3DGS). Recently, there has been a surge in NeRF-based SLAM, while 3DGS-based SLAM is sparse. A novel 3DGS-based SLAM approach with a fusion of deep visual feature, dual keyframe selection and 3DGS is presented in this paper. Compared with the existing methods, the proposed tracking is achieved by feature extraction and motion filter on each frame. The joint optimization of poses and 3D Gaussians runs through the entire mapping process. Additionally, the coarse-to-fine pose estimation and compact Gaussian scene representation are implemented by dual keyframe selection and novel loss functions. Experimental results demonstrate that the proposed algorithm not only outperforms the existing methods in tracking and mapping, but also has less memory usage.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

### Problems Addressed by the Paper

This paper proposes a novel SLAM system based on 3D Gaussian Splatting (3DGS), named MotionGS. Specifically, the paper addresses the following issues:

1. **High-Fidelity Scene Reconstruction**: Traditional dense visual SLAM methods struggle to achieve high-fidelity scene representation and face difficulties in reconstructing fine textures and repetitive scenes. MotionGS achieves high-precision real-time tracking and reconstruction by integrating deep visual features, a dual keyframe selection strategy, and 3DGS technology.

2. **Memory Optimization**: Existing NeRF-based methods rely on ray-tracing volume rendering, which is time-consuming and unreliable; while 3DGS-based methods render quickly, there is still room for improvement in memory usage. MotionGS reduces memory consumption by introducing new loss functions and masking mechanisms, thereby decreasing the number of Gaussian points involved in optimization.

3. **Real-Time Performance Enhancement**: By designing a new dual keyframe strategy (motion keyframes and information keyframes), MotionGS not only improves tracking accuracy but also optimizes real-time rendering effects. Experimental results show that this method outperforms existing methods on the Replica and TUM-RGBD datasets in terms of tracking and mapping, achieving the current best performance with a running speed of 2.5 frames per second.

In summary, this paper aims to develop a SLAM system capable of real-time high-precision localization and map construction while optimizing memory usage efficiency.