Dynamic SLAM With 3-D Gaussian Splatting Supporting Monocular Sensing

Monocular simultaneous localization and mapping (SLAM) systems have garnered significant attention due to their cost effectiveness and adaptability. However, they face substantial challenges in dense reconstruction tasks within dynamic environments, primarily due to the lack of reliable depth information. This article presents a real-time dynamic SLAM system based on 3-D Gaussian splatting (3DGS) that achieves robust pose estimation and high-fidelity dense reconstruction using only monocular camera input. The system effectively segments dynamic objects by integrating feature point tracking with a lightweight depth estimation network. We innovatively propose a pixel-set completion strategy that employs semantic cues and depth estimation to perform the cluster compensation for anomalous optical flow pixels, enhancing the system stability while maintaining real-time performance. Furthermore, we design a Gaussian dense initialization method based on inactive feature points and introduce dynamic pixel-set masking with the residual loss for soft suppression, significantly improving reconstruction quality in dynamic scenes. Experimental results demonstrate that our method achieves reconstruction quality comparable with RGB-D systems without relying on accurate depth values, while offering superior adaptability and practical utility.