DyGS-SLAM: Dynamic Object Segmentation and Multiview Long-Term Gaussian Consistency for High-Fidelity SLAM

Robust and accurate pose estimation in dynamic environments remains a fundamental challenge for visual simultaneous localization and mapping (SLAM). Recent advances in integrating 3D Gaussian Splatting (3DGS) into SLAM have demonstrated its potential for achieving high-quality scene reconstruction with explicit Gaussian representations. Compared with LiDAR-based systems, monocular and RGB-D cameras are lightweight, inexpensive, and capable of providing rich visual and geometric information, making them highly attractive for practical applications such as human–robot interaction and autonomous driving. However, visual SLAM systems are often sensitive to illumination variations, occlusions, and dynamic objects, which introduce unstable features that may cause pose drift and pseudo-static artifacts, thereby compromising localization stability and map integrity. To address these challenges, this paper presents DyGS-SLAM, a 3DGS-based visual SLAM framework designed for dynamic environments and compatible with both monocular and RGB-D sensors. The proposed system integrates vision-based object detection with depth sensing to perform coarse-to-fine dynamic segmentation, while a temporal-consistency refinement module leverages long-term multi-view observations to suppress residual dynamic artifacts. Extensive experiments on the TUM RGB-D and KITTI datasets show that DyGS-SLAM achieves improved localization accuracy compared with recent dynamic-aware baselines, reducing ATE by 4.7% on TUM and 6.3% on KITTI, while running in real time at 13.8 FPS with 7.4 GB GPU memory on an RTX 4080 Ti. These results highlight the potential of low-cost visual sensors for constructing high-quality 3D maps, enabling practical deployment in a wide range of intelligent sensing and robotic applications.