Gaussian Splatting SLAM Based on Loop Closure Pose Optimization and Comprehensive Loss Function

With their high-fidelity scene representation capabilities, Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have deeply attracted attention in Simultaneous Localization and Mapping (SLAM) community. This paper proposes a 3DGS-based SLAM method that integrates the loop closure detection module and designs a comprehensive loss function. Graphical optimizations are used to optimize trace threads and correct for accumulated errors. The experimental results show that the average improvement is about 54% compared with SplaTAM in terms of tracking, and the image similarity is also improved in terms of mapping.